Prediction system, prediction method, prediction device, learning device, program, and trained model

Abstract

To provide a prediction system, prediction method, prediction device, learning device, program, and trained model for appropriately predicting future falls.SOLUTION: A prediction system provided herein comprises: a model generation phase for storing associated information in which detection information obtained by sensing at least a part of the body and prediction information regarding future falls are associated with each other; and a prediction phase for predicting information regarding future falls of a subject based on the associated information upon receipt of an input of the detection information on the subject.SELECTED DRAWING: Figure 29

Description

[Technical Field]

[0001] The present invention relates to a prediction system, a prediction method, a prediction device, a learning device, a program, and a trained model. [Background technology]

[0002] A system is known that measures data on the foot pressure distribution of a pedestrian, analyzes the risk of falling, and presents the risk (see, for example, Patent Document 1). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Laid-Open No. 2002-345786 Summary of the Invention [Problem to be solved by the invention]

[0004] However, the conventional technology disclosed in Patent Document 1 only presents a rough indication of the current risk of falling (high / low), and it is difficult to predict the probability of falling in the future.

[0005] In view of the above circumstances, an object of the present invention is to provide a technology for suitably predicting future falls. [Means for solving the problem]

[0006] One aspect of the present invention is a prediction system that includes a correspondence information storage unit that stores correspondence information in which detection information sensed from at least a part of the body is pre-associated with prediction information regarding future falls, and a prediction unit that, when the detection information of a subject is input, predicts information regarding the subject's future falls based on the correspondence information.

[0007] Another aspect of the present invention is a prediction method in a prediction system, which includes a correspondence information storage process in which a correspondence information storage unit stores correspondence information in which detection information obtained by sensing at least a part of the body is pre-associated with prediction information regarding future falls, and a prediction process in which a prediction unit, when the detection information of a subject is input, predicts information regarding the subject's future falls based on the correspondence information.

[0008] Another aspect of the present invention is a prediction device that includes a correspondence information reference unit that references correspondence information in which detection information sensed from at least a part of the body is associated in advance with prediction information regarding future falls, and a prediction unit that, when the detection information of a subject is input, predicts information regarding the subject's future falls based on the correspondence information.

[0009] Another aspect of the present invention is a learning device comprising: a learning information acquisition unit that acquires learning information in which detection information obtained by sensing at least a part of the body and prediction information regarding future falls are associated for each of a plurality of subjects; and a learning unit that performs machine learning based on the learning information.

[0010] Another aspect of the present invention is a program for causing a computer to execute a correspondence information reference means for referencing correspondence information in which detection information obtained by sensing at least a part of the body is pre-matched with prediction information regarding future falls, and a prediction means for predicting information regarding a future fall of a subject based on the correspondence information when the detection information of the subject is input.

[0011] Another aspect of the present invention is a program for causing a computer to execute a learning information acquisition means for acquiring learning information in which detection information obtained by sensing at least a part of the body and prediction information regarding future falls are associated for each of a plurality of subjects, and a learning means for performing machine learning based on the learning information.

[0012] Another aspect of the present invention is a trained model in which detection information sensed from at least a part of the body and prediction information regarding future falls are generated by machine learning based on learning information associated with each of a plurality of subjects, and when the detection information of a subject is input, the trained model predicts information regarding the subject's future falls. [Effects of the Invention]

[0013] Future falls can be predicted in an advantageous manner. [Brief explanation of the drawings]

[0014] [Figure 1] FIG. 1 is a conceptual diagram for explaining an outline of a first embodiment (the same applies to second to fourth embodiments). [Figure 2] FIG. 2 is an explanatory diagram illustrating sensor information and the like. [Figure 3] FIG. 1 illustrates an example of the configuration of a prediction device according to a first embodiment. [Figure 4] FIG. 1 is a diagram illustrating a trained model in the first embodiment. [Figure 5] 4 is a flowchart showing an example of the operation of the prediction device in the first embodiment. [Figure 6] FIG. 1 is a diagram illustrating an example of the configuration of a model generating device in a first embodiment. [Figure 7] FIG. 2 is a diagram illustrating a training dataset in the first embodiment. [Figure 8] 4 is a flowchart showing an example of the operation of the model generating device in the first embodiment. [Figure 9] FIG. 4 is a diagram illustrating another example of the configuration of the model generating device in the first embodiment. [Figure 10] FIG. 3 is a diagram illustrating an input data set in the first embodiment. [Figure 11] FIG. 10 is a diagram illustrating another trained model in the first embodiment. [Figure 12] FIG. 10 is a diagram illustrating another training data set in the first embodiment. [Figure 13] FIG. 10 is a diagram illustrating an example of the configuration of a prediction device according to a second embodiment. [Figure 14] FIG. 10 is a diagram illustrating a trained model in the second embodiment. [Figure 15] 10 is a flowchart showing an example of the operation of the prediction device in the second embodiment. [Figure 16] FIG. 10 is a diagram illustrating an example of the configuration of a model generating device in a second embodiment. [Figure 17] FIG. 10 is a diagram illustrating a training data set in the second embodiment. [Figure 18] 10 is a flowchart showing an example of the operation of the model generating device in the second embodiment. [Figure 19] FIG. 10 is a diagram illustrating another exemplary configuration of a model generating device in the second embodiment. [Figure 20] FIG. 10 is a diagram illustrating an input data set in the second embodiment. [Figure 21] FIG. 10 is a diagram illustrating another trained model in the second embodiment. [Figure 22] FIG. 10 is a diagram illustrating another training data set in the second embodiment. [Figure 23] FIG. 10 is a diagram illustrating an example of the configuration of a prediction device according to a third embodiment. [Figure 24] FIG. 11 is a diagram illustrating a trained model in the third embodiment. [Figure 25] 11 is a flowchart showing an example of the operation of the prediction device in the third embodiment. [Figure 26] FIG. 10 is a diagram illustrating an example of the configuration of a model generating device in a third embodiment. [Figure 27] FIG. 11 is a diagram illustrating a training data set in the third embodiment. [Figure 28] 10 is a flowchart showing an example of the operation of the model generating device in the third embodiment. [Figure 29] FIG. 10 is a conceptual diagram for explaining an outline of a fifth embodiment (similar to the sixth and seventh embodiments). [Figure 30]FIG. 10 is a diagram illustrating an example of the configuration of a prediction device according to a fifth embodiment. [Figure 31] FIG. 13 is a diagram illustrating a first trained model in the fifth embodiment. [Figure 32] FIG. 13 is a diagram illustrating a second trained model in the fifth embodiment. [Figure 33] 13 is a flowchart showing an example of the operation of the prediction device in the fifth embodiment. [Figure 34] FIG. 12 is a diagram illustrating an example of the configuration of a first trained model generation device in a fifth embodiment. [Figure 35] FIG. 13 is a diagram illustrating a training dataset for a first trained model in the fifth embodiment. [Figure 36] 10 is a flowchart showing an example of operation of the first trained model generation device in the fifth embodiment. [Figure 37] FIG. 13 is a diagram illustrating another example configuration of the first trained model generation device in the fifth embodiment. [Figure 38] FIG. 13 is a diagram illustrating an input data set in the fifth embodiment. [Figure 39] FIG. 11 is a diagram illustrating another trained model for the first trained model in the fifth embodiment. [Figure 40] FIG. 13 is a diagram illustrating another training data set for the first trained model in the fifth embodiment. [Figure 41] FIG. 13 is a diagram illustrating an example of the configuration of a second trained model generation device in the fifth embodiment. [Figure 42] FIG. 13 is a diagram illustrating a training dataset for a second trained model in the fifth embodiment. [Figure 43] 10 is a flowchart showing an example of operation of the second trained model generation device in the fifth embodiment. [Figure 44] FIG. 13 is a diagram illustrating an example of the configuration of a prediction device according to a sixth embodiment. [Figure 45] FIG. 20 is a diagram illustrating a first trained model in the sixth embodiment. [Figure 46]13 is a flowchart showing an example of the operation of the prediction device in the sixth embodiment. [Figure 47] 13 is a flowchart showing an example of the operation of the second trained model generation device in the sixth embodiment. [Figure 48] FIG. 13 is a diagram illustrating a training dataset for a second trained model in the fifth embodiment. [Figure 49] 10 is a flowchart showing an example of operation of the second trained model generation device in the fifth embodiment. [Figure 50] FIG. 13 is a diagram illustrating an example of the configuration of a prediction device according to a seventh embodiment. [Figure 51] FIG. 20 is a diagram illustrating a first trained model in the seventh embodiment. [Figure 52] 13 is a flowchart showing an example of the operation of the prediction device in the seventh embodiment. [Figure 53] 13 is a flowchart showing an example of operation of the first trained model generation device in the seventh embodiment. [Figure 54] FIG. 13 is a diagram illustrating a training dataset for a first trained model in the seventh embodiment. [Figure 55] 13 is a flowchart showing an example of operation of the first trained model generation device in the seventh embodiment. [Figure 56] FIG. 10 is an explanatory diagram illustrating evaluation of a model. [Figure 57] FIG. 10 is an explanatory diagram illustrating evaluation of a model. DETAILED DESCRIPTION OF THE INVENTION

[0015] An outline of the first embodiment (similar to the second to fourth embodiments) of the present invention will be described with reference to Figures 1 and 2. Figure 1 is a conceptual diagram for explaining the outline of the first embodiment (similar to the second to fourth embodiments).

[0016] As shown in Fig. 1, the system according to the first embodiment (as well as the second to fourth embodiments) includes a model generation phase and a prediction phase. The model generation phase is a phase in which a trained model to be used in the prediction phase is generated. The prediction phase is a phase in which a fall probability is predicted using the trained model generated in the model generation phase.

[0017] In the following description of each embodiment (first to seventh embodiments), a person who collects data for generating a trained model in the model generation phase may be referred to as a subject, and a person who predicts the probability of a fall using the trained model in the prediction phase may be referred to as a target. In other words, in the model generation phase, data is collected from the subject to generate a trained model, and in the prediction phase, the probability of a fall is predicted for the target.

[0018] The system according to the first embodiment (as well as the second to fourth embodiments) uses at least sensor information (details to be described later) that senses walking to predict the probability of falling within a predetermined period (for example, within one year) after sensing. Furthermore, the system according to the first embodiment (as well as the second to fourth embodiments) uses at least the predicted result of the fall probability (0% to 100%) to determine whether intervention by a specialist is necessary to deal with future falls. Intervention by a specialist may include individual diagnosis in person or remotely, exercise advice, etc.

[0019] In predicting the probability of a fall and determining whether intervention is necessary, diagnostic information (details will be described later) and questionnaire information (details will be described later) may be used in addition to sensor information. In other words, the system according to the first embodiment (as well as the second to fourth embodiments) may include at least one of the following configurations 1 to 9.

[0020] (Configuration 1) A configuration that predicts the probability of falling based on sensor information and determines whether intervention is necessary based on the predicted result of the probability of falling. (Configuration 2) A configuration in which the probability of falling is predicted based on sensor information, and whether intervention is necessary is determined based on the predicted result of the probability of falling and the diagnostic information. (Configuration 3) A configuration in which the probability of falling is predicted based on sensor information, and whether intervention is necessary is determined based on the predicted result of the probability of falling and questionnaire information. (Configuration 4) A configuration in which the probability of falling is predicted based on sensor information, and whether intervention is necessary is determined based on the predicted result of the probability of falling, diagnosis information, and questionnaire information. (Configuration 5) A configuration that predicts the probability of a fall based on sensor information and diagnostic information, and determines whether intervention is necessary based on the predicted result of the fall probability. (Configuration 6) A configuration in which the probability of a fall is predicted based on sensor information and diagnostic information, and whether or not intervention is necessary is determined based on the predicted result of the probability of a fall and questionnaire information. (Configuration 7) A configuration that predicts the probability of falling based on sensor information and questionnaire information, and determines whether intervention is necessary based on the predicted result of the probability of falling. (Configuration 8) A configuration that predicts the probability of falling based on sensor information and questionnaire information, and determines whether intervention is necessary based on the predicted result of the probability of falling and diagnostic information. (Configuration 9) A configuration in which the probability of a fall is predicted based on sensor information, diagnosis information, and questionnaire information, and whether or not intervention is necessary is determined based on the predicted result of the probability of a fall.

[0021] Machine learning and statistical methods are used to predict the probability of a fall. Therefore, prior to predicting the probability of a fall, necessary information is collected and a trained model (trained model) or statistical model is created. As an example, in the case of configuration 1 (the same applies to configurations 2 to 4), sensor information from each of a large number of subjects and fall occurrence information (fall history information) indicating whether or not each subject actually fell within a predetermined period (within one year) from the date the sensor information was acquired are collected, and a trained model is created that uses feature information based on the sensor information (or sensor information) as an input sample and the fall history information as an output sample, or a statistical model that uses feature information based on the sensor information (or sensor information) as an explanatory variable and the fall history information as a target variable.

[0022] (sensor information) FIG. 2 is an explanatory diagram illustrating sensor information, etc. FIG. 2(A) is an example of a sensor shoe (gait data acquisition device). As shown in FIG. 2(A), the sensor shoe includes a metatarsophalangeal joint flexion sensor (reference symbol A in FIG. 2(A)), a talocrural flexion sensor (reference symbol B in FIG. 2(A)), an arch flexion sensor (reference symbol C in FIG. 2(A)), a talus flexion sensor (not shown), a heel pressure sensor (reference symbol D in FIG. 2(A)), a metatarsophalangeal joint pressure sensor (reference symbol E in FIG. 2(A)), a toe pressure sensor (reference symbol F in FIG. 2(A)), and a box (reference symbol G in FIG. 2(A)). The box contains an acceleration sensor (reference symbol H in FIG. 2(A)), a data processing unit (reference symbol I in FIG. 2(A)), and a power supply unit (reference symbol J in FIG. 2(A)). The data processing unit includes a microcomputer and a data output unit.

[0023] The metatarsophalangeal joint flexion sensor measures the degree of flexion around the metatarsophalangeal joint axis of the foot. The talocrural flexion sensor measures the degree of flexion around the talocrural joint axis of the ankle. The arch flexion sensor measures the degree of flexion of the arch. The talus flexion sensor measures the degree of flexion around the subtalar joint axis of the ankle. The heel pressure sensor measures pressure at the heel on the sole of the foot. The metatarsophalangeal joint pressure sensor measures pressure at the metatarsophalangeal joint on the sole of the foot. The toe pressure sensor measures pressure at the big toe on the sole of the foot. The acceleration sensor is configured to be able to measure acceleration in three axes, and measures the up-down acceleration, front-back acceleration, and left-right acceleration (inward and outward directions of the foot) of the lower leg. These are output by the data output unit.

[0024] Alternatively, a sensor may be used that measures foot extension during walking instead of or in addition to foot flexion during walking. The metatarsophalangeal joint flexion sensor may calculate the degree of extension from the degree of flexion.

[0025] Hereinafter, the sensor information acquired from the sensor shoes (for example, all or part of the information acquired from each of the above-mentioned sensors) may be referred to as sensor information 1 to sensor information N. Each of the multiple output values ​​(sensor values) from one sensor may also be referred to as sensor information. For example, the output value at each of multiple timings (for example, each of ten equal timings) within one walking cycle from a metatarsophalangeal joint bending sensor may also be referred to as sensor information.

[0026] The sensor information (sensor information 1 to sensor information N) is stored in association with the person performing the gait measurement (the subject in the model generation phase, and the target person in the prediction phase). For example, the sensor information of person A is stored in association with identification information that identifies person A, and the sensor information of person B is stored in association with identification information that identifies person B.

[0027] (feature information) Various feature amounts relating to walking can be extracted from the sensor information 1 to sensor information N. For example, the following feature amount information can be acquired. [Basic data] Step count, walking pace (steps / min) Step length Walking speed [Walking Cycle] Heel strike and toe off time Average walking cycle time (seconds) Stance phase, swing phase (time, %) Single leg support period (time, %) Double support period (time, %) [Waveform (time normalized)] -One walking cycle, arithmetic average waveform, standard deviation waveform [Various representative values] Locker area Maximum, minimum, and range values ​​(width) during stance and swing phases Unit time (slope), average value of the above data

[0028] Hereinafter, each piece of feature amount information (all or part of each piece of feature amount information described above) may be referred to as feature amount information 1 to feature amount information N. Note that the "N" in feature amount information N and the "N" in sensor information N do not indicate the same number. In other words, the "N" in sensor information N indicates one or more pieces of sensor information acquired from the sensor shoes, and the "N" in feature amount information N indicates one or more pieces of feature amount information extracted from one or more pieces of sensor information, and does not indicate whether the two pieces of feature amount information match in number.

[0029] The feature information (feature information 1 to feature information N) may be stored in association with the person taking the gait measurement (the subject in the model generation phase, and the target person in the prediction phase). For example, when feature information (feature information 1 to feature information N) is extracted from the sensor information of person A (sensor information stored in association with identification information identifying person A), the feature information may be stored in association with person A (stored in association with the identification information identifying person A), and when feature information (feature information 1 to feature information N) is extracted from the sensor information of person B (sensor information stored in association with identification information identifying person B), the feature information may be stored in association with person B (stored in association with the identification information identifying person B).

[0030] Fig. 2(B) shows an example of sensor information acquired from the sensor shoes. Fig. 2(B) shows, as an example, sensor information of the walking cycle acquired from a toe pressure sensor, sensor information of the walking cycle acquired from a metatarsophalangeal joint bending sensor, and sensor information of the walking cycle acquired from an acceleration sensor.

[0031] Regarding walking, instead of or in addition to sensing the legs using sensor shoes, sensors may be worn on the arms or head to sense the arms or head. In other words, the sensor information does not necessarily have to be information obtained from sensor shoes.

[0032] (Survey information) The survey information may be created based on the following survey, for example. 1. Attributes and basic information (age, gender, height, weight, foot length, etc.) Measure as necessary. Blood pressure, BMI (obesity level information), bone density, etc. may also be included. 2. Fall history (1) Whether or not you have fallen within the past three months If you fell, please provide the specific details (month (morning, afternoon, evening, night), location, and circumstances (e.g., tripped over a step during a morning walk. Only a scratch, did not go to the hospital, etc.)) (2) Whether or not you have fallen within the past year (excluding the past three months) If there is a fall, the same as (1). 3. Condition of legs and hips (Locomotive Syndrome 5 Survey) (1) How difficult is it to go up and down stairs? (2) How difficult is it to walk briskly? (3) How long can you walk without stopping? (4) How difficult is it to carry about 2 kg of groceries home? (5) How difficult is it to do moderately heavy household chores (such as using a vacuum cleaner or putting away or putting away futons)? 4. Physical condition (1) Subjective symptoms of decreased motor function (check the appropriate box) Motor function information that indicates subjective symptoms of motor function decline. For example, decreased muscle strength (upper limbs, trunk, lower limbs), decreased motor speed (slow movements), decreased reaction time (longer reaction times), decreased manual dexterity, decreased sense of balance, decreased physical flexibility, etc.

[0033] Furthermore, instead of or in addition to the above, as questionnaire information, musculoskeletal evaluation information evaluated by responses to multiple questions may be used as a method of evaluating a person's functions of standing, walking, or sitting, which are related to the physical movements necessary for daily life.

[0034] The questionnaire information may also include at least one of information on past medical conditions, information on experiences with abnormalities in bones, joints, or muscles, information on care services, information on medication, and information on exercise habits.

[0035] Hereinafter, the information obtained based on the above-mentioned questionnaire (all or part of the above-mentioned question items) may be referred to as questionnaire information 1 to questionnaire information N. Note that the "N" in questionnaire information N indicates the number of diagnostic information items (number of items) used, and is not particularly related to the "N" in sensor information N, etc., or the "N" in feature amount information N.

[0036] The questionnaire information (questionnaire information 1 to questionnaire information N) is stored in association with the person who performed the gait measurement (the subject in the model generation phase, and the target person in the prediction phase). For example, the questionnaire information of person A is stored in association with the identification information that identifies person A, and the questionnaire information of person B is stored in association with the identification information that identifies person B.

[0037] (Diagnostic information) The diagnostic information is gait evaluation information in which an expert assessor evaluates the gait abnormality by referring to the gait of the measurer (the subject in the model generation phase, and the target in the prediction phase). For example, the GARS-M (Modified Gait Abnormality Scale, hereinafter referred to as GARS) may be used as the diagnostic information. GARS scores each of the following seven items on a scale of 0-3, and evaluates the gait abnormality with an overall evaluation value (0-21 points). GARS1:variability GARS2: Lack of momentum GARS3: Stagger GARS4: Foot contact GARS5: Hip joint range of motion GARS6: Backward arm swing GARS7: Synchrony of arm swing and heel strike

[0038] The expert may check the walking behavior of the person being measured (the test subject in the model generation phase, and the target person in the prediction phase) in captured images (videos) of the person's walking behavior and assign a score to each of the seven items above. It is preferable that the captured images are taken on the day of the walking measurement. The images may be taken together with the walking measurement or separately from the walking measurement (in other words, the expert may check the person's appearance with the sensor shoes on and assign a score, or check the person's appearance without the sensor shoes on and assign a score), but the conditions for whether the sensor shoes are on or off should be the same.

[0039] Hereinafter, information obtained based on the above-mentioned expert diagnosis (for example, scores for all or some of the seven GARS items (GARS1 to GARS7)) may be referred to as diagnostic information 1 to diagnostic information N. For example, when all seven GARS items are used as GARS diagnostic information, the information becomes diagnostic information 1 to diagnostic information 7. Note that the "N" in diagnostic information N indicates the number of diagnostic information items (number of items) used, and is not particularly related to the "N" in sensor information N, etc., the "N" in feature amount information N, or the "N" in questionnaire information N.

[0040] The diagnostic information (diagnostic information 1 to diagnostic information N) is stored in association with the person who performed the gait measurement. For example, the diagnostic information of person A is stored in association with identification information that identifies person A, and the diagnostic information of person B is stored in association with identification information that identifies person B.

[0041] (Fall history information) Fall history information (results of whether or not a fall occurred) is acquired and fall history information is accumulated. For example, one year after the date of walking measurement (measurement using sensor shoes), a hearing may be conducted regarding whether or not a fall occurred during the one-year period from the walking measurement date to the day one year after the walking measurement date (for example, responses may be collected as a questionnaire at a later date). The hearing regarding whether or not a fall occurred may be conducted several times during the one-year period. For example, three months after the walking measurement date, a hearing regarding whether or not a fall occurred during the three-month period from the walking measurement date to the day three months after the walking measurement date; six months after the walking measurement date, a hearing regarding whether or not a fall occurred during the three-month period from the day three months after the walking measurement date to the day six months after the walking measurement date; nine months after the walking measurement date, a hearing regarding whether or not a fall occurred during the three-month period from the day six months after the walking measurement date to the day nine months after the walking measurement date; and one year after the walking measurement date, a hearing regarding whether or not a fall occurred during the three-month period from the day nine months after the walking measurement date to the day one year after the walking measurement date.

[0042] The fall history information is stored in association with the person performing the gait measurement (the subject in the model generation phase, and the target person in the prediction phase). For example, the fall history information of person A is stored in association with identification information that identifies person A, and the fall history information of person B is stored in association with identification information that identifies person B.

[0043] (First embodiment) A first embodiment of the present invention will be described below with reference to Figs. 3 to 12. The first embodiment corresponds to configurations 1 to 4 among configurations 1 to 9 described in Fig. 1. Although it has been explained that the sensor information does not necessarily have to be information obtained from a sensor shoe, in the first embodiment (and similarly in other embodiments), the sensor information will be described as information obtained from a sensor shoe.

[0044] (Prediction device 10) The prediction device 10 predicts the probability of a subject falling using a trained model 2100 (described later). The trained model is a learning model represented by parameters (trained parameters) optimized through training (learning). The trained model is generated by training the learning model and optimizing the parameters through training. The prediction device 10 also determines whether intervention is necessary using at least the prediction result of the fall probability. The prediction device 10 may be, for example, a personal computer.

[0045] FIG. 3 is a diagram illustrating an example of the configuration of a prediction device 10. As illustrated in FIG. 3, the prediction device 10 includes an acquisition unit 100, a feature extraction unit 101, a fall prediction unit (estimation unit) 102, an intervention necessity determination unit 103, an output unit 104, a sensor information storage unit 105, a feature information storage unit 108, a model storage unit 109, a prediction result storage unit 110, and a determination result storage unit 111. The units within the prediction device 10 (as well as the units within the prediction device 12, the prediction device 14, the prediction device 16, the prediction device 18, the prediction device 20, and the prediction device 22) can communicate with each other via a bus. Although not illustrated, the prediction device 10 (as well as the units within the prediction device 12, the prediction device 14, the prediction device 16, the prediction device 18, the prediction device 20, and the prediction device 22) includes an operation unit that receives operations from an operator. Although indicated by the dashed lines, the prediction device 10 may further include a diagnosis information storage unit 106 and a questionnaire information storage unit 107.

[0046] The feature amount estimation unit 101, fall prediction unit 102, and intervention necessity determination unit 103 are realized by a CPU or the like. Each storage unit (sensor information storage unit 105, diagnosis information storage unit 106, questionnaire information storage unit 107, feature amount information storage unit 108, model storage unit 109, prediction result storage unit 110, and determination result storage unit 111) is realized by an HDD, an SSD (Solid State Drive), or the like. The output unit 104 may be, for example, a display unit such as a liquid crystal display, an audio output unit such as a speaker, a communication unit capable of transmitting information to the outside, or a printing unit that prints and issues paper media or the like.

[0047] The acquisition unit 100 acquires sensor information. In this example, the acquisition unit 100 acquires sensor information (pressure sensor values, bending sensor values, acceleration sensor values) obtained from the sensor shoe. The acquisition unit 100 may acquire the sensor information directly from the sensor shoe (for example, by wireless communication), or may acquire the sensor information via another device (a device that acquired the sensor information from the sensor shoe) (for example, by wireless communication, wired communication, wireless communication, or using a storage medium). After acquiring the sensor information, the acquisition unit 100 stores the sensor information in the sensor information storage unit 105.

[0048] The sensor information storage unit 105 stores the sensor information acquired by the acquisition unit 100. Specifically, the sensor information storage unit 105 stores the sensor information for each subject.

[0049] Furthermore, the acquiring unit 100 may acquire diagnostic information of a diagnosis made by an expert on the subject. When the acquiring unit 100 acquires the diagnostic information, it stores the diagnostic information in the diagnostic information storage unit 106. The diagnostic information storage unit 106 stores the diagnostic information acquired by the acquiring unit 100. Specifically, the diagnostic information storage unit 106 stores the diagnostic information for each subject. Note that in the first embodiment, the acquiring unit 100 may acquire diagnostic information, but may not acquire diagnostic information. Therefore, if the acquiring unit 100 does not acquire diagnostic information, the diagnostic information storage unit 106 may not exist.

[0050] Furthermore, the acquisition unit 100 may acquire survey information for a survey to which the subject himself / herself has answered. When the acquisition unit 100 acquires the survey information, it stores the survey information in the survey information storage unit 107. The survey information storage unit 107 stores the survey information acquired by the acquisition unit 100. Specifically, the survey information storage unit 107 stores the survey information for each subject. Note that in the first embodiment, the acquisition unit 100 may acquire the survey information, but may not acquire the survey information. Therefore, if the acquisition unit 100 does not acquire the survey information, the survey information storage unit 107 may not exist.

[0051] The feature extraction unit 101 extracts feature amounts from the sensor information stored in the sensor information storage unit 105. After extracting the feature amounts, the feature extraction unit 101 stores feature amount information indicating the extracted feature amounts in the feature amount information storage unit 108.

[0052] The feature amount information storage unit 108 stores feature amount information indicating the feature amounts extracted by the feature amount extraction unit 101. Specifically, the feature amount information storage unit 108 stores feature amount information for each subject.

[0053] The model storage unit 109 stores the trained model 2100. The trained model 2100 is generated in a model generation device 30 (described later) and supplied to the prediction device 10. The trained model 2100 generated in the model generation device 30 is supplied to the prediction device 10 from the model generation device 30 via communication or using a storage medium (such as a USB memory), and is stored in the model storage unit 109. The same applies to the supply and storage of other trained models.

[0054] FIG. 4 is a diagram illustrating a trained model 2100. The trained model 2100 is, for example, a model represented using a neural network as shown in FIG. 4. The trained model 2100 may be, for example, a DNN (Deep Neural Network). The trained parameters include, for example, the number of layers of the neural network, the number of neurons in each layer, the connection relationships between neurons, the weights (connection loads) of the connections between each neuron, and the thresholds of each neuron. The trained model 2100 shown in FIG. 4 includes an input layer 2101, one or more intermediate layers 2102 (hidden layers), and an output layer 2103. Each of the layers 2101, 2102, and 2103 includes one or more neurons. The number of intermediate layers 2101, the number of neurons in each layer, etc. are set by the trained parameters.

[0055] The input layer 2101 receives as input the feature information stored in the feature information storage unit 108. The output layer 303 outputs an output value indicating the probability of falling within a predetermined period (specifically, within one year) and an output value indicating the probability of not falling within the predetermined period. In other words, the trained model 2100 is a trained model that, when the feature information of a subject (feature information 1 to feature information N) is input to the input layer 2101, outputs from the output layer 2103 the probability that the subject will fall within the predetermined period and the probability that the subject will not fall within the predetermined period.

[0056] The fall prediction unit 102 predicts the probability that the subject will fall within a predetermined period in the future (specifically, within one year) using the feature amount information stored in the feature amount information storage unit 108 and the trained model 2100 stored in the model storage unit 109. Having predicted the probability of falling, the fall prediction unit 102 stores the prediction result (a value between 0% and 100%) in the prediction result storage unit 110.

[0057] The prediction result storage unit 110 stores the prediction result by the fall prediction unit 102. Specifically, the prediction result storage unit 110 stores the prediction result for each subject.

[0058] The intervention necessity determination unit 103 determines whether intervention by an expert is necessary based on the prediction result stored in the prediction result storage unit 110. For example, the intervention necessity determination unit 103 may determine that intervention by an expert is unnecessary when the prediction result is less than a predetermined threshold (e.g., 60%), and may determine that intervention by an expert is necessary when the prediction result is equal to or greater than the predetermined threshold. Having determined whether intervention is necessary, the intervention necessity determination unit 103 stores the determination result (necessary / not necessary) in the determination result storage unit 111.

[0059] Note that a threshold value of 60% for comparison with the prediction results stored in the prediction result storage unit 110 may be stored in a location that can be referenced by the intervention necessity determination unit 103, or a table in which a value of a fall probability less than 60% corresponds to an indication that intervention is not required and a value of a fall probability of 60% or more corresponds to an indication that intervention is required may be stored. In short, the intervention necessity determination unit 103 may refer to the prediction results stored in the prediction result storage unit 110 and the threshold value (or the table) and determine whether intervention by a specialist is required. The same applies to the intervention necessity determination units of other embodiments.

[0060] Furthermore, the intervention necessity determination unit 103 may determine whether or not intervention by a specialist is necessary based on one or both of the diagnostic information (the diagnostic information stored in the diagnostic information storage unit 106) and the questionnaire information (the questionnaire information stored in the questionnaire information storage unit 107) in addition to the prediction result. When the diagnostic information or the questionnaire information is used to determine whether or not intervention is necessary, each of them may be scored (assigned a point). The diagnostic information may be scored, for example, using an overall evaluation value of the diagnostic information 1 to N (for example, an overall evaluation value (0 to 21 points) of GARS1 to GARS7). The questionnaire information may be scored, for example, by weighting and adding the answers to the questionnaire information 1 to N (each of the questionnaire question items). Note that in order to match the prediction result, which is the probability of falling (0 to 100%), the diagnostic information and the questionnaire information are scored so that the smaller the value, the less intervention is required.

[0061] When determining whether or not intervention is necessary based on the prediction result and the diagnosis information, the intervention necessity determination unit 103 determines whether or not intervention is necessary by using a determination score Y1 AB Calculate the judgment score Y1 AB is the predetermined reference value R1 AB If the score is less than Y1, intervention is deemed unnecessary. AB is the predetermined reference value R1 AB If the above is true, intervention may be deemed necessary. A is the predicted value, X B is the score value of the diagnostic information, and a and b are coefficients.

[0062] Y1 AB = a × X1 A +b×X B … (1)

[0063] When determining whether or not intervention is necessary based on the prediction result and the questionnaire information, the intervention necessity determination unit 103 determines whether or not intervention is necessary by using the following formula (2): AC Calculate the judgment score Y1 AC is the predetermined reference value R1 AC If the score is less than Y1, intervention is deemed unnecessary. AC is the predetermined reference value R1 AC If the above is true, intervention may be deemed necessary. A is the predicted value, X C is the score value of the survey information, and c and d are coefficients.

[0064] Y1 AC =c×X1 A +d×X C … (2)

[0065] When determining whether or not intervention is necessary based on the prediction result, the diagnosis information, and the questionnaire information, the intervention necessity determination unit 103 determines whether or not intervention is necessary by using the following formula (3): ABC Calculate the judgment score Y1 ABC is the predetermined reference value R1 ABC If the score is less than Y1, intervention is deemed unnecessary. ABC is the predetermined reference value R1 ABC If the above is true, intervention may be deemed necessary. A is the predicted value, X B is the score value of the diagnostic information, X C is the score value of the survey information, and e, f, and g are coefficients.

[0066] Y1 ABC =e×X1 A +f×X B +g×X C … (3)

[0067] The output unit 104 outputs the prediction result by the fall prediction unit 102 (the prediction result stored in the prediction result storage unit 110) and the determination result by the intervention necessity determination unit 103 (the determination result stored in the determination result storage unit 111). Specifically, the output unit 104 outputs the prediction result and the determination result for each subject. Note that, instead of or in addition to the above, the output unit 104 may transmit each piece of information to the subject's personal device (smartphone, personal computer). For example, if the determination result indicates that intervention is necessary, the output unit 104 may transmit screen information including the prediction result, the determination result (necessary), and information regarding an appointment for diagnosis (intervention) to the subject's personal device.

[0068] Note that the prediction results (possibility of falling in the future) may be classified in advance according to value (for example, less than 20%, 20-40%, 40-60%, 60-80%, 80-100%), and information according to the classification (no problem, ..., very dangerous) may be stored, and the output unit 104 (the same applies to the output units in other embodiments) may output output information according to the classification based on the prediction results. The same applies to the output units in other embodiments.

[0069] 5 is a flowchart showing an example of the operation of the prediction device 10. Specifically, the flowchart in FIG. 5 shows an example of the operation of the acquisition unit 100, the feature extraction unit 101, the fall prediction unit 102, the intervention necessity determination unit 103, and the output unit 104.

[0070] The prediction device 10 (acquisition unit 100) acquires sensor information (step S10). The prediction device 10 (acquisition unit 100) stores the sensor information (stores it in the sensor information storage unit 105). Subsequently, the prediction device 10 (feature extraction unit 101) extracts features (step S11). The prediction device 10 (feature extraction unit 101) stores the feature information (stores it in the feature information storage unit 108). Subsequently, the prediction device 10 (fall prediction unit 102) inputs the feature information into the trained model 2100 to predict the probability of a fall (step S12). The prediction device 10 (fall prediction unit 102) stores the prediction result (stores it in the prediction result storage unit 110). Subsequently, the prediction device 10 (intervention necessity determination unit 103) determines whether intervention by an expert is necessary (step S13). The prediction device 10 (intervention necessity determination unit 103) stores the determination result (stores it in the prediction result storage unit 110). Subsequently, the prediction device 10 (output unit 104) outputs the prediction result by the fall prediction unit 102 and the determination result by the intervention necessity determination unit 103 (step S14). Then, this flowchart ends.

[0071] (Model generation device 30) The model generation device 30 generates a trained model 2100 to be used in the prediction device 10. Specifically, the model generation device 30 trains a trained model using a training dataset 1100 (described later) derived from multiple measurers (subjects) who acquired sensor information, and generates the trained model 2100. The model generation device 30 may be, for example, a personal computer.

[0072] FIG. 6 is a diagram illustrating an example configuration of the model generation device 30. As illustrated in FIG. 6, the model generation device 30 includes an acquisition unit 300, a model generation unit 302, and a storage unit 309. The units within the model generation device 30 (including the units of the model generation device 32, the unit of the model generation device 34, the unit of the model generation device 36, the unit of the first trained model generation device 38, the unit of the second trained model generation device 39, the unit of the second trained model generation device 41, and the unit of the first trained model generation device 42) can communicate with each other via a bus. Although not illustrated, the model generation device 30 (including the model generation device 32, the model generation device 34, the first trained model generation device 38, the second trained model generation device 39, the second trained model generation device 41, and the first trained model generation device 42) also includes an operation unit that accepts operations from an operator and a communication unit that communicates with other devices. The model generation unit 302 is implemented by a CPU or the like. The storage unit 309 is realized by an HDD, an SSD (Solid State Drive), or the like.

[0073] The acquisition unit 300 acquires the training dataset 1100 from outside. The acquisition unit 300 may acquire the training dataset 1100 from outside using a storage medium (such as a USB memory), or may acquire the training dataset 1100 from outside via communication. The same applies to the supply and storage of other training datasets. Having acquired the training dataset 1100, the acquisition unit 300 stores the training dataset 1100 in the storage unit 2309.

[0074] The training dataset 1100 (as well as other training datasets) is composed of multiple training samples (training sample 1 to training sample N). Each training sample is information obtained from a respective subject. For example, training sample 1 is information obtained from subject 1, and training sample 2 is information obtained from subject 2. Each training sample includes an input sample and an output sample. The input sample is data input to the input layer when training the training model. The output sample is data (also called teacher data or correct answer label) that serves as the correct answer for comparison with the output value from the output layer when training the training model.

[0075] FIG. 7 is a diagram illustrating a training dataset 1100. The acquiring unit 300 acquires, for example, the training dataset 1100 shown in FIG. 7. The acquiring unit 300 may acquire training samples 1 to N collectively or individually. The acquiring unit 300 may also acquire feature information and performance information individually. In this case, for example, common identification information (such as a subject number) may be added to each of the feature information and performance information so that a single training sample can be generated from the feature information and performance information after acquisition (i.e., so that it can be recognized that they constitute the same training sample (i.e., information on the subject)).

[0076] 7 includes, as an input sample (symbol A), feature information based on the sensor information obtained from each subject, and as an output sample (symbol B), actual information on whether or not each subject has fallen within a predetermined period (specifically, within one year from the date of acquisition of each sensor information) (fall history survey information).

[0077] Note that feature information corresponding to the input sample is obtained from sensor information, but the step of obtaining (extracting) the feature information from the sensor information is performed in a device (not shown; a training dataset generation device including a feature extraction unit similar to the feature extraction unit 101 of the prediction device 10 shown in FIG. 3) different from the model generation device 30. In other words, a device (training dataset generation device) different from the model generation device 30 generates the training dataset 1100, and the model generation device 30 acquires the training dataset 1100 from the training dataset generation device.

[0078] The storage unit 309 stores the learning data set 1100 (learning sample 1 to learning sample N in FIG. 7) acquired by the acquisition unit 300. Furthermore, the storage unit 309 stores a learning model and parameters.

[0079] The model generation unit 302 trains a learning model using the learning dataset 1100 stored in the storage unit 309, and generates a trained model 2100 (details will be described later). The model generation unit 302 stores the generated trained model 2100 in the storage unit 309.

[0080] Specifically, the model generation unit 302 inputs each of the input samples (feature information of each of the training samples 1 to N in FIG. 7) to the input layer 2101, updates the parameters of the training model so as to reduce the difference between each of the output values ​​obtained from the output layer 2103 and each of the output samples (performance information of each of the training samples 1 to N in FIG. 7), and generates the trained model 2100. As an example, the model generation unit 302 may generate the trained model 2100 using backpropagation, as shown in the following (1) to (5).

[0081] (1) The model generation unit 302 inputs input samples (feature information of learning samples 1 to N) of the learning dataset 1100 stored in the storage unit 309 to the input layer 2101, and performs calculation processing in the forward propagation direction of the learning model to obtain output values ​​from the output layer 2103. The output values ​​from the output layer 2103 are values ​​that represent the respective probabilities of falling / not falling (likelihoods). (2) The model generation unit 302 calculates the error between the output value from the output layer 2103 and the output sample for all input samples, for example, by backpropagation. The output sample is actual information (whether or not a fall actually occurred). For example, when the output sample of learning sample 1 indicates no fall, if the input sample (feature information) of learning sample 1 is input to the input layer 2101 and the output value obtained from the output layer 2103 indicates a 0% probability of fall (a 100% probability of not falling), then the error for learning sample 1 is zero. Also, when the output sample of learning sample 2 indicates a fall, if the input sample (feature information) of learning sample 2 is input to the input layer 2101 and the output value obtained from the output layer 2103 indicates a 100% probability of fall (a 0% probability of not falling), then the error for learning sample 2 is zero. (3) The model generation unit 302 determines whether the calculated error is within a predetermined value. (4) If the calculated error is within a predetermined value, the model generation unit 302 determines that various parameters such as the connection weights between neurons (such as the connection weights between neurons and the thresholds of each neuron) have been optimized, and ends learning (training). In other words, the generation of the trained model 2100 by the model generation unit 302 is completed (the trained model in which the various parameters are trained parameters is referred to as the trained model 2100). (5) If the calculated error is not within a predetermined value, the model generation unit 302 updates various parameters based on the calculated errors. Thereafter, the model generation unit 302 inputs the input sample again to the input layer 2101 of the learning model whose parameters have been updated, and repeatedly updates various parameters until the error falls within a predetermined value.

[0082] As described above, the trained model 2100 is a learning model that has been trained and generated so that when feature information (feature information 1 to feature information N) of each subject is input to the input layer 2101, the fall probability of each subject (the probability of falling within a specified period, the probability of not falling within a specified period) output from the output layer 2103 matches (as closely as possible) the actual fall history of each subject (the result of whether or not a fall occurred).

[0083] Fig. 8 is a flowchart showing an example of the operation of the model generation device 30. Specifically, the flowchart in Fig. 8 shows an example of the operation of the model generation unit 302. It is assumed that at the start of the flowchart in Fig. 8, the learning dataset 1100, learning model, and parameters as shown in Fig. 7 are stored in the storage unit 309.

[0084] The model generation device 30 (model generation unit 302; the same applies hereinafter) inputs input samples (feature information) of each subject (each subject) into the learning model and acquires each output value from the learning model (step S20). Next, the model generation device 30 calculates the error between each output value and each output sample (whether or not a fall occurred) (step S21). Next, the model generation device 30 determines whether the error is within a predetermined value (step S22). If the error is not within the predetermined value (step S22: NO), the model generation device 30 updates various parameters based on each error (step S23) and returns to step S20. That is, steps S20 to S23 are repeatedly executed until the error is within the predetermined value. If the error is within the predetermined value (step S22: YES), the model generation device 30 ends training (step S24). That is, the generation of the trained model 2100 by the model generation device 30 is completed (the trained model in which the various parameters are trained parameters is the trained model 2100). Then, this flowchart ends.

[0085] (Modification of the first embodiment) The following describes modifications of the first embodiment. Each modification may be combined as appropriate.

[0086] (Modification 1 of the first embodiment) In the first embodiment, it has been explained that the process of obtaining feature information from sensor information is performed in a device (learning dataset generation device) different from the model generation device 30, but the model generation device 30 may be responsible for the process of obtaining feature information from sensor information.

[0087] 9 is a diagram illustrating another exemplary configuration of the model generation device 30. As illustrated in FIG. 9, the model generation device 30 may further include an input data storage unit 305 and a feature extraction unit 301 in addition to the acquisition unit 300, the model generation unit 302, and the storage unit 309. The feature extraction unit 301 extracts features from sensor information, similar to the feature extraction unit 101 of the prediction device 10.

[0088] Fig. 10 is a diagram illustrating an input data set 1110. The acquisition unit 300 of the model generating device 30 shown in Fig. 6 acquires the learning data set 1100 shown in Fig. 7 and stores it in the storage unit 309, while the acquisition unit 300 of the model generating device 30 shown in Fig. 9 acquires the input data set 1110 shown in Fig. 10 and stores it in the input data storage unit 305.

[0089] The input data set 1110 (as well as the other input data sets) is composed of information for multiple training samples (information for training sample 1 to information for training sample N). The information for each training sample is information obtained from a respective subject. For example, the information for training sample 1 is information obtained from subject 1, and the information for training sample 2 is information obtained from subject 2.

[0090] The input dataset 1110 (information for each learning sample) shown in Fig. 10 differs from the learning dataset 1100 (each learning sample) shown in Fig. 7 in the following respects: The learning dataset 1100 shown in Fig. 7 includes feature information, which is an input sample (symbol A), and performance information, which is an output sample (symbol B), whereas the input dataset 1110 shown in Fig. 10 includes sensor information, which is information (symbol C) for obtaining the input sample (feature information), and performance information, which is an output sample (symbol B).

[0091] The acquiring unit 300 of the model generating device 30 shown in FIG. 9 acquires an input data set 1110. The acquiring unit 300 may acquire information for learning sample 1 to information for learning sample N collectively or individually. The acquiring unit 300 may also acquire sensor information and performance information individually. In this case, for example, common identification information (such as a subject number) may be added to each of the sensor information and performance information so that information for one learning sample can be generated from the sensor information and performance information after acquisition (i.e., so that it can be recognized that they constitute information for the same learning sample (i.e., information for one subject)).

[0092] The feature extraction unit 301 extracts features from the sensor information in the input dataset 1110 (the feature extraction unit 301 extracts feature information 1 to feature information N to be input samples of training sample 1 based on sensor information 1 to sensor information N included in the information for training sample 1, extracts feature information 1 to feature information N to be input samples of training sample 2 based on sensor information 1 to sensor information N included in the information for training sample 2, ... extracts feature information 1 to feature information N to be input samples of training sample N based on sensor information 1 to sensor information N included in the information for training sample N), thereby generating a training dataset 1100. The feature extraction unit 301 stores the generated training dataset 1100 in the storage unit 309. Hereinafter, the model generation device 30 shown in FIG. 9 generates a trained model 2100 using the training dataset 1100, similar to the model generation device 30 shown in FIG. 6.

[0093] (Modification 2 of the first embodiment) Fig. 11 is a diagram illustrating another trained model 2120. The trained model 2100 shown in Fig. 4 inputs feature information and outputs the probability of falling (probability of not falling), but a trained model 2120 as shown in Fig. 11 may also be used, which inputs sensor information and outputs the probability of falling (probability of not falling).

[0094] Fig. 12 is a diagram illustrating another training dataset 1120. When using the trained model 2100 shown in Fig. 4, the training dataset 1100 shown in Fig. 7 (or the input dataset 1110 shown in Fig. 10) is used, but when using the trained model 2120 shown in Fig. 11, the training dataset 1120 shown in Fig. 12 is used.

[0095] The training dataset 1120 (respective training samples) shown in Fig. 12 differs from the training dataset 1100 (respective training samples) shown in Fig. 7 in the following respects: The training dataset 1100 shown in Fig. 7 includes feature information, which is an input sample (symbol A), and performance information, which is an output sample (symbol B), whereas the training dataset 1120 shown in Fig. 12 includes sensor information, which is an input sample (symbol A), and performance information, which is an output sample (symbol B). Note that when the trained model 2120 or the training dataset 1120 is used, feature information is not used, and therefore, needless to say, no configuration related to features is required in each device (prediction device, model generation device).

[0096] (Second embodiment) Next, a second embodiment of the present invention will be described with reference to Figures 13 to 22. The second embodiment corresponds to configurations 5 and 6 among configurations 1 to 9 described in Figure 1. Note that the second embodiment will be described mainly focusing on the parts that are different from the first embodiment, and the description of the parts that are common to the first embodiment will be omitted in part or in whole.

[0097] (Prediction Device 12) The prediction device 12 predicts the probability of a subject falling using a trained model 2200 (described later). The prediction device 12 also determines whether intervention is necessary using at least the prediction result of the fall probability.

[0098] Fig. 13 is a diagram showing an example configuration of prediction device 12. As shown in Fig. 13, prediction device 12 includes an acquisition unit 120, a feature extraction unit 121, a fall prediction unit (estimation unit) 122, an intervention necessity determination unit 123, an output unit 124, a sensor information storage unit 125, a diagnosis information storage unit 126, a feature information storage unit 128, a model storage unit 129, a prediction result storage unit 130, and a determination result storage unit 131. Note that, although indicated by a dashed line, prediction device 12 may further include a questionnaire information storage unit 127.

[0099] The feature extraction unit 121, output unit 124, sensor information storage unit 125, diagnostic information storage unit 126, feature information storage unit 128, prediction result storage unit 130, and judgment result storage unit 131 provided in the prediction device 12 of the second embodiment are similar to the feature extraction unit 101, output unit 104, sensor information storage unit 105, diagnostic information storage unit 106, feature information storage unit 108, prediction result storage unit 110, and judgment result storage unit 111 provided in the prediction device 10 of the first embodiment.

[0100] The acquisition unit 120, like the acquisition unit 100 of the prediction device 10 of the first embodiment, acquires sensor information obtained from the sensor shoes and stores it in the sensor information storage unit 125. The acquisition unit 120 also acquires diagnostic information of a diagnosis made by an expert on the subject and stores it in the diagnostic information storage unit 126.

[0101] Furthermore, similar to the acquisition unit 100 of the prediction device 10 of the first embodiment, the acquisition unit 120 may acquire survey information of a survey answered by the subject himself / herself and store it in the survey information storage unit 127. If the acquisition unit 120 does not acquire survey information, the survey information storage unit 127 may not exist.

[0102] The model storage unit 129 stores the trained model 2200. The trained model 2200 is generated by a model generation device 32 (described later) and supplied to the prediction device 12.

[0103] Fig. 14 is a diagram illustrating a trained model 2200. The trained model 2200 is, for example, a model expressed using a neural network as shown in Fig. 14. As shown in Fig. 14, the trained model 2200 of the second embodiment is a trained model that, when feature information (feature information 1 to feature information N) and diagnostic information (diagnostic information 1 to diagnostic information N) of a subject are input to an input layer 2201, outputs from an output layer 2203 the probability that the subject will fall within a predetermined period and the probability that the subject will not fall within a predetermined period.

[0104] The fall prediction unit 122 predicts the probability that the subject will fall within a specified period of time in the future using the feature information stored in the feature information storage unit 128, the diagnostic information stored in the diagnostic information storage unit 126, and the trained model 2200 stored in the model storage unit 129, and stores the prediction result in the prediction result storage unit 130.

[0105] The intervention necessity determination unit 123 determines whether intervention by an expert is necessary based on the prediction result stored in the prediction result storage unit 130. For example, the intervention necessity determination unit 123 may determine that intervention by an expert is unnecessary when the prediction result is less than a predetermined threshold (e.g., 60%), and may determine that intervention by an expert is necessary when the prediction result is equal to or greater than the predetermined threshold. Having determined whether intervention is necessary, the intervention necessity determination unit 123 stores the determination result (necessary / not necessary) in the determination result storage unit 131.

[0106] Furthermore, the intervention necessity determination unit 123 may determine whether or not intervention by a specialist is necessary based on questionnaire information (questionnaire information stored in the questionnaire information storage unit 127) in addition to the prediction result. When the questionnaire information is used to determine whether or not intervention is necessary, the questionnaire information may be scored. For example, the questionnaire information may be scored by weighting and adding up the answers to questionnaire information 1 to questionnaire information N (each of the questions in the questionnaire). Note that in order to match the prediction result, which is the probability of falling (0 to 100%), the questionnaire information is scored so that the smaller the value, the less intervention is required.

[0107] When determining whether or not intervention is necessary based on the prediction result and the questionnaire information, the intervention necessity determination unit 123 determines whether or not intervention is necessary by using the following formula (4): AC Calculate the judgment score Y2 AC is the predetermined reference value R2 AC If the score is less than Y2, intervention is deemed unnecessary. AC is the predetermined reference value R2 AC If the condition is more than X2, intervention may be deemed necessary. A is the predicted value, X Cis the score value of the survey information, and h and i are coefficients.

[0108] Y2 AC =h×X2 A +i×X C … (4)

[0109] Fig. 15 is a flowchart showing an example of the operation of the prediction device 12. Specifically, the flowchart in Fig. 15 shows an example of the operation of the acquisition unit 120, the feature extraction unit 121, the fall prediction unit 122, the intervention necessity determination unit 123, and the output unit 124.

[0110] The prediction device 12 (acquisition unit 120) acquires sensor information and diagnostic information (step S30). The prediction device 12 (acquisition unit 120) stores the sensor information and diagnostic information (sensor information is stored in the sensor information storage unit 125, and diagnostic information is stored in the diagnostic information storage unit 126). Subsequently, the prediction device 12 (feature extraction unit 121) extracts features (step S31). The prediction device 12 (feature extraction unit 121) stores the feature information (stores it in the feature information storage unit 128). Subsequently, the prediction device 12 (fall prediction unit 122) inputs the feature information and diagnostic information into the trained model 2200 to predict the probability of a fall (step S32). The prediction device 12 (fall prediction unit 122) stores the prediction result (stores it in the prediction result storage unit 130). Subsequently, the prediction device 12 (intervention necessity determination unit 123) determines whether intervention by a specialist is necessary (step S33). The prediction device 12 (intervention necessity determination unit 123) stores the determination result (stores it in the prediction result storage unit 130). Subsequently, the prediction device 12 (output unit 124) outputs the prediction result by the fall prediction unit 122 and the determination result by the intervention necessity determination unit 123 (step S34). Then, this flowchart ends.

[0111] (Model Generator 32) The model generation device 32 generates a trained model 2200 to be used in the prediction device 12. The model generation device 32 trains the trained model using a training dataset 1200 (described later) derived from multiple measurers (subjects) who acquired sensor information, and generates the trained model 2200.

[0112] 16 is a diagram showing an example of the configuration of the model generation device 32. As shown in FIG.

[0113] The acquiring unit 320 externally acquires the training dataset 1200. After acquiring the training dataset 1200, the acquiring unit 320 stores the training dataset 1200 in the storage unit 2329.

[0114] FIG. 17 is a diagram illustrating a training dataset 1200. The acquiring unit 320 acquires, for example, the training dataset 1200 as shown in FIG. 17. The acquiring unit 320 may acquire training samples 1 to N collectively or individually. The acquiring unit 320 may also acquire feature information, diagnostic information, and performance information individually. In this case, for example, common identification information (such as a subject number) may be added to each of the feature information, diagnostic information, and performance information so that one training sample can be generated from the feature information, diagnostic information, and performance information after acquisition (i.e., so that it can be recognized that they constitute the same training sample (i.e., information on the subject)).

[0115] 17 includes, as an input sample (symbol A), feature information based on the sensor information obtained from each subject, and actual diagnostic information by an expert for each subject. Also, as an output sample (symbol B), it includes actual information (fall history survey information) on whether each subject has fallen within a predetermined period (specifically, within one year from the date of acquisition of each sensor information).

[0116] Note that feature information corresponding to a part of the input samples is obtained from the sensor information, but the step of obtaining the feature information from the sensor information (extracting step) is performed in a device (not shown; a training dataset generation device including a feature extraction unit similar to the feature extraction unit 121 of the prediction device 12 shown in FIG. 13) different from the model generation device 32. In other words, a device (training dataset generation device) different from the model generation device 32 generates the training dataset 1200, and the model generation device 32 acquires the training dataset 1200 from the training dataset generation device.

[0117] The storage unit 329 stores the learning data set 1200 (learning sample 1 to learning sample N in FIG. 17) acquired by the acquisition unit 320. Furthermore, the storage unit 329 stores a learning model and parameters.

[0118] The model generation unit 322 trains a learning model using the learning dataset 1200 stored in the storage unit 329, and generates a trained model 2200. The model generation unit 322 stores the generated trained model 2200 in the storage unit 329.

[0119] Specifically, the model generation unit 322 inputs each of the input samples (the feature information and diagnostic information of each of the training samples 1 to N in FIG. 17) to the input layer 2201, updates the parameters of the training model so as to reduce the difference between each of the output values ​​obtained from the output layer 2203 and each of the output samples (the performance information of each of the training samples 1 to N in FIG. 17), and generates the trained model 2200. For example, the model generation unit 322 may generate the trained model 2200 using an error backpropagation algorithm, similar to the model generation unit 302 of the model generation device 30 of the first embodiment.

[0120] As described above, trained model 2200 is a learning model that has been trained and generated so that when feature information (feature information 1 to feature information N) and diagnostic information (diagnostic information 1 to diagnostic information N) of each subject are input to input layer 2201, the fall probability of each subject (probability of falling within a specified period, probability of not falling within a specified period) output from output layer 2203 matches (as closely as possible) each subject's actual fall history (results of whether or not a fall occurred).

[0121] Fig. 18 is a flowchart showing an example of the operation of the model generation device 32. Specifically, the flowchart in Fig. 18 shows an example of the operation of the model generation unit 322. It is assumed that at the start of the flowchart in Fig. 18, the learning dataset 1200, learning model, and parameters as shown in Fig. 17 are stored in the storage unit 329.

[0122] The model generation device 32 (model generation unit 322; the same applies hereinafter) inputs input samples (feature information, diagnosis information) of each subject (each subject) into the learning model and acquires each output value from the learning model (step S40). Next, the model generation device 32 calculates the error between each output value and each output sample (presence or absence of fall) (step S41). Next, the model generation device 32 determines whether the error is within a predetermined value (step S42). If the error is not within the predetermined value (step S42: NO), the model generation device 32 updates various parameters based on each error (step S43) and returns to step S40. That is, steps S40 to S43 are repeatedly executed until the error is within the predetermined value. If the error is within the predetermined value (step S42: YES), the model generation device 32 ends training (step S44). That is, the generation of the trained model 2200 by the model generation device 32 is completed (the trained model in which the various parameters are trained parameters is the trained model 2200). Then, this flowchart ends.

[0123] (Modification of the second embodiment) The following describes modifications of the second embodiment. The modifications may be combined as appropriate.

[0124] (Modification 1 of the second embodiment) In the second embodiment, it has been explained that the process of obtaining feature information from sensor information is performed in a device (learning dataset generation device) different from the model generation device 32, but the model generation device 32 may be responsible for the process of obtaining feature information from sensor information.

[0125] 19 is a diagram illustrating another exemplary configuration of the model generation device 32. As illustrated in FIG. 19, the model generation device 32 may further include an input data storage unit 325 and a feature extraction unit 321 in addition to the acquisition unit 320, the model generation unit 322, and the storage unit 329. The feature extraction unit 321 extracts features from sensor information, similar to the feature extraction unit 121 of the prediction device 12.

[0126] Fig. 20 is a diagram illustrating an input data set 1210. The acquisition unit 320 of the model generation device 32 shown in Fig. 16 acquires the learning data set 1200 as shown in Fig. 17 and stores it in the storage unit 329, while the acquisition unit 320 of the model generation device 32 shown in Fig. 19 acquires the input data set 1210 as shown in Fig. 20 and stores it in the input data storage unit 325.

[0127] The input dataset 1210 (information for each learning sample) shown in Fig. 20 differs from the learning dataset 1200 (each learning sample) shown in Fig. 17 in the following respects: The learning dataset 1200 shown in Fig. 17 includes feature information and diagnostic information that are input samples (symbol A), and performance information that are output samples (symbol B), whereas the input dataset 1210 shown in Fig. 20 includes sensor information that is information (symbol C) for obtaining the input samples (feature information), diagnostic information that is the input samples (symbol A), and performance information that is the output samples (symbol B).

[0128] The acquiring unit 320 of the model generating device 32 shown in FIG. 19 acquires the input data set 1210. The acquiring unit 320 may acquire the information for learning sample 1 to information for learning sample N collectively or individually. The acquiring unit 320 may also acquire the sensor information, the diagnostic information, and the performance information individually. In this case, for example, common identification information (such as a subject number) may be added to each of the sensor information, the diagnostic information, and the performance information so that information for one learning sample can be generated from the sensor information, the diagnostic information, and the performance information after acquisition (i.e., so that it can be recognized that they constitute information for the same learning sample (i.e., information for one subject)).

[0129] The feature extraction unit 321 extracts features from the sensor information in the input dataset 1210 and generates a training dataset 1200. The feature extraction unit 321 stores the generated training dataset 1200 in the storage unit 329. Thereafter, the model generation device 32 shown in FIG. 19 generates a trained model 2200 using the training dataset 1200, similar to the model generation device 32 shown in FIG. 16.

[0130] (Modification 2 of the second embodiment) Fig. 21 is a diagram illustrating another trained model 2220. The trained model 2200 shown in Fig. 14 inputs feature information and diagnostic information and outputs the probability of falling (probability of not falling), but a trained model 2220 as shown in Fig. 14 inputs sensor information and diagnostic information and outputs the probability of falling (probability of not falling) may also be used.

[0131] Fig. 22 is a diagram illustrating another training dataset 1220. When using a trained model 2200 such as that shown in Fig. 14, the training dataset 1200 shown in Fig. 17 (or the input dataset 1210 shown in Fig. 20) is used, but when using a trained model 2220 such as that shown in Fig. 21, the training dataset 1220 shown in Fig. 22 is used.

[0132] The training dataset 1220 (respective training samples) shown in Fig. 22 differs from the training dataset 1200 (respective training samples) shown in Fig. 17 in the following respects: The training dataset 1200 shown in Fig. 17 includes feature information and diagnostic information, which are input samples (symbol A), and performance information, which are output samples (symbol B), whereas the training dataset 1220 shown in Fig. 22 includes sensor information and diagnostic information, which are input samples (symbol A), and performance information, which are output samples (symbol B). Note that when the trained model 2220 or the training dataset 1220 is used, feature information is not used, and therefore, needless to say, no configuration related to features is required in each device (prediction device, model generation device).

[0133] (Third embodiment) Next, a third embodiment will be described with reference to Figures 23 to 28. The third embodiment corresponds to configuration 9 among configurations 1 to 9 described in Figure 1. Note that the third embodiment will be described mainly focusing on the parts that are different from the first and second embodiments, and the description of the parts that are common to the first and second embodiments will be omitted in part or in whole.

[0134] (Prediction Device 14) The prediction device 14 predicts the probability of a subject falling using a trained model 2300 (described later). The prediction device 14 also determines whether intervention is necessary using the predicted result of the fall probability.

[0135] Fig. 23 is a diagram showing an example configuration of prediction device 14. As shown in Fig. 23, prediction device 14 includes an acquisition unit 140, a feature extraction unit 141, a fall prediction unit (estimation unit) 142, an intervention necessity determination unit 143, an output unit 144, a sensor information storage unit 145, a diagnosis information storage unit 126, a questionnaire information storage unit 147, a feature information storage unit 148, a model storage unit 149, a prediction result storage unit 150, and a determination result storage unit 151.

[0136] The feature extraction unit 141, output unit 144, sensor information storage unit 145, diagnostic information storage unit 146, questionnaire information storage unit 147, feature information storage unit 148, prediction result storage unit 150, and judgment result storage unit 151 provided in the prediction device 14 of the third embodiment are similar to the feature extraction unit 101, output unit 104, sensor information storage unit 105, diagnostic information storage unit 106, questionnaire information storage unit 107, feature information storage unit 108, prediction result storage unit 110, and judgment result storage unit 111 provided in the prediction device 10 of the first embodiment.

[0137] The acquisition unit 140, like the acquisition unit 100 of the prediction device 10 of the first embodiment, acquires sensor information obtained from the sensor shoes and stores it in the sensor information storage unit 145. Similarly to the acquisition unit 120 of the prediction device 12 of the second embodiment, the acquisition unit 140 acquires diagnostic information of a diagnosis made by an expert on the subject and stores it in the diagnostic information storage unit 146. The acquisition unit 140 also acquires questionnaire information of a questionnaire answered by the subject himself / herself and stores it in the questionnaire information storage unit 147.

[0138] The model storage unit 149 stores the trained model 2300. The trained model 2300 is generated by a model generation device 34 (described later) and supplied to the prediction device 14.

[0139] Fig. 24 is a diagram illustrating a trained model 2300. The trained model 2300 is, for example, a model expressed using a neural network as shown in Fig. 24. As shown in Fig. 24, the trained model 2300 of the third embodiment is a trained model that, when feature information (feature information 1 to feature information N), diagnosis information (diagnosis information 1 to diagnosis information N), and questionnaire information (questionnaire information 1 to questionnaire information N) of a subject are input to an input layer 2301, outputs from an output layer 2303 the probability that the subject will fall within a predetermined period and the probability that the subject will not fall within a predetermined period.

[0140] The fall prediction unit 142 predicts the probability that the subject will fall within a specified future period using the feature information stored in the feature information storage unit 148, the diagnostic information stored in the diagnostic information storage unit 146, the diagnostic information stored in the questionnaire information storage unit 147, and the trained model 2300 stored in the model storage unit 149, and stores the prediction result in the prediction result storage unit 150.

[0141] The intervention necessity determination unit 143 determines whether intervention by an expert is necessary based on the prediction result stored in the prediction result storage unit 150. For example, the intervention necessity determination unit 143 may determine that intervention by an expert is unnecessary when the prediction result is less than a predetermined threshold (e.g., 60%), and may determine that intervention by an expert is necessary when the prediction result is equal to or greater than the predetermined threshold. Having determined whether intervention is necessary, the intervention necessity determination unit 143 stores the determination result (necessary / not necessary) in the determination result storage unit 151.

[0142] Fig. 25 is a flowchart showing an example of the operation of prediction device 14. Specifically, the flowchart in Fig. 25 shows an example of the operation of acquisition unit 140, feature amount extraction unit 141, fall prediction unit 142, intervention necessity determination unit 143, and output unit 144.

[0143] The prediction device 14 (acquisition unit 140) acquires sensor information, diagnosis information, and questionnaire information (step S50). The prediction device 14 (acquisition unit 140) stores the sensor information, diagnosis information, and questionnaire information (sensor information is stored in the sensor information storage unit 145, diagnosis information is stored in the diagnosis information storage unit 146, and questionnaire information is stored in the questionnaire information storage unit 147). Subsequently, the prediction device 14 (feature extraction unit 141) extracts features (step S51). The prediction device 14 (feature extraction unit 141) stores the feature information (stores it in the feature information storage unit 148). Subsequently, the prediction device 14 (fall prediction unit 142) inputs the feature information, diagnosis information, and questionnaire information into the trained model 2300 to predict the probability of a fall (step S52). The prediction device 14 (fall prediction unit 142) stores the prediction result (stores it in the prediction result storage unit 150). Next, the prediction device 14 (intervention necessity determination unit 143) determines whether intervention by an expert is necessary (step S53). The prediction device 14 (intervention necessity determination unit 143) stores the determination result (stores it in the prediction result storage unit 150). Next, the prediction device 14 (output unit 144) outputs the prediction result by the fall prediction unit 142 and the determination result by the intervention necessity determination unit 143 (step S54). Then, this flowchart ends.

[0144] (Model Generator 34) The model generation device 34 generates a trained model 2300 to be used in the prediction device 14. The model generation device 34 trains the learning model using a learning dataset 1300 (described later) derived from multiple measurers (subjects) who acquired sensor information, and generates the trained model 2300.

[0145] 26 is a diagram showing an example of the configuration of the model generation device 34. As shown in FIG. 26, the model generation device 34 includes an acquisition unit 340, a model generation unit 342, and a storage unit 349.

[0146] The acquiring unit 340 externally acquires the training dataset 1300. After acquiring the training dataset 1300, the acquiring unit 340 stores the training dataset 1300 in the storage unit 2349.

[0147] FIG. 27 is a diagram illustrating a training dataset 1300. The acquiring unit 340 acquires, for example, the training dataset 1300 as shown in FIG. 27. The acquiring unit 340 may acquire training samples 1 to N collectively or individually. The acquiring unit 340 may also acquire feature information, diagnostic information, questionnaire information, and performance information individually. In this case, for example, common identification information (such as a subject number) may be added to each of the feature information, diagnostic information, questionnaire information, and performance information so that one training sample can be generated from the feature information, diagnostic information, questionnaire information, and performance information after acquisition (i.e., so that it can be recognized that they constitute the same training sample (i.e., information on the subject)).

[0148] 27 includes, as an input sample (symbol A), feature information based on the sensor information obtained from each subject, actual diagnostic information by an expert for each subject, and questionnaire information obtained from each subject. Also, as an output sample (symbol B), it includes actual information (fall history survey information) on whether each subject has fallen within a predetermined period (specifically, within one year from the date of acquisition of each sensor information).

[0149] Note that feature information corresponding to a part of the input sample is obtained from the sensor information, but the step of obtaining the feature information from the sensor information (extracting step) is performed in a device (not shown; a training dataset generation device including a feature extraction unit similar to the feature extraction unit 141 of the prediction device 14 shown in FIG. 23) different from the model generation device 34. In other words, a device (training dataset generation device) different from the model generation device 34 generates the training dataset 1300, and the model generation device 34 acquires the training dataset 1300 from the training dataset generation device.

[0150] The storage unit 349 stores the learning data set 1300 (learning sample 1 to learning sample N in FIG. 27) acquired by the acquisition unit 340. Furthermore, the storage unit 349 stores a learning model and parameters.

[0151] The model generation unit 342 trains a learning model using the learning dataset 1300 stored in the storage unit 349, and generates a trained model 2300. The model generation unit 342 stores the generated trained model 2300 in the storage unit 349.

[0152] Specifically, the model generation unit 342 inputs each of the input samples (the feature information, diagnostic information, and questionnaire information of each of the training samples 1 to N in FIG. 27) to the input layer 2301, updates the parameters of the training model so as to reduce the difference between each of the output values ​​obtained from the output layer 2303 and each of the output samples (the performance information of each of the training samples 1 to N in FIG. 27), and generates the trained model 2300. For example, the model generation unit 342 may generate the trained model 2300 using an error backpropagation algorithm, similar to the model generation unit 302 of the model generation device 30 of the first embodiment.

[0153] As described above, the trained model 2300 is a learning model that has been trained and generated so that when the feature information (feature information 1 to feature information N), diagnostic information (diagnostic information 1 to diagnostic information N), and questionnaire information (questionnaire information 1 to questionnaire information N) of each subject are input to the input layer 2301, the fall probability of each subject (the probability of falling within a specified period, the probability of not falling within a specified period) output from the output layer 2303 matches (as closely as possible) the actual fall history of each subject (the result of whether or not a fall occurred).

[0154] Fig. 28 is a flowchart showing an example of the operation of the model generation device 34. Specifically, the flowchart in Fig. 28 shows an example of the operation of the model generation unit 342. It is assumed that at the start of the flowchart in Fig. 28, the learning dataset 1300, learning model, and parameters as shown in Fig. 27 are stored in the storage unit 349.

[0155] The model generation device 34 (model generation unit 342; the same applies below) inputs input samples (feature information, diagnosis information, questionnaire information) of each subject into the learning model and acquires each output value from the learning model (step S60). Next, the model generation device 34 calculates the error between each output value and each output sample (whether or not a fall occurred) (step S61). Next, the model generation device 34 determines whether the error is within a predetermined value (step S62). If the error is not within the predetermined value (step S62: NO), the model generation device 34 updates various parameters based on each error (step S63) and returns to step S60. That is, steps S60 to S63 are repeatedly executed until the error is within the predetermined value. If the error is within the predetermined value (step S62: YES), the model generation device 34 ends training (step S64). That is, the generation of the trained model 2300 by the model generation device 34 is completed (the trained model in which the various parameters are trained parameters is the trained model 2300). Then, this flowchart ends.

[0156] The third embodiment also has the same modifications (Modification 1 and Modification 2) as the first and second embodiments, but the description thereof will be omitted.

[0157] (Fourth embodiment) Next, a fourth embodiment will be described. The fourth embodiment corresponds to configurations 7 and 8 among configurations 1 to 9 described in Fig. 1. Note that the fourth embodiment will be described mainly focusing on the parts that are different from the first to third embodiments, and the description of the parts that are common to the first to third embodiments will be omitted in part or in whole.

[0158] The fourth embodiment replaces the diagnostic information in the second embodiment with questionnaire information, and therefore a brief explanation will be omitted. That is, the prediction device (prediction device 12) of the second embodiment predicts the probability of a subject falling within a predetermined future period using feature information, diagnostic information, and a trained model (trained model 2200) and determines whether or not intervention by a specialist is required based on the prediction result (or the prediction result and questionnaire information), whereas the prediction device (prediction device 16, not shown) of the fourth embodiment predicts the probability of a subject falling within a predetermined future period using feature information, questionnaire information, and a trained model (trained model 2400, not shown) and determines whether or not intervention by a specialist is required based on the prediction result (or the prediction result and diagnostic information).

[0159] The trained model 2400 (not shown) of the fourth embodiment is a trained model that, when feature information (feature information 1 to feature information N) of a subject and questionnaire information (questionnaire information 1 to questionnaire information N) are input to the input layer 2401, outputs the probability that the subject will fall within a specified period and the probability that the subject will not fall within a specified period from the output layer 2403.

[0160] A model generation device (model generation device 36, not shown) of the fourth embodiment generates a trained model 2400 (not shown) using a training dataset 1400 (not shown). The training dataset 1400 (each training sample) includes, as input samples, feature information based on each sensor information obtained from each subject, and questionnaire information obtained from each subject. Also, as output samples, it includes actual information (fall history investigation information) on whether each subject has fallen within a predetermined period (specifically, within one year from the date of acquisition of each sensor information).

[0161] A model generation unit 362 (not shown) of the model generation device 36 inputs each input sample to the input layer 2401, updates the parameters of the learning model so as to reduce the difference between each output value obtained from the output layer 2403 and the output samples, and generates a trained model 2400. For example, the model generation unit 362 may generate the trained model 2400 using the backpropagation algorithm, similar to the model generation unit 302 of the model generation device 30 of the first embodiment.

[0162] As described above, the trained model 2400 is a learning model that has been trained and generated so that when the feature information (feature information 1 to feature information N) and questionnaire information (questionnaire information 1 to questionnaire information N) of each subject are input to the input layer 2401, the fall probability of each subject (the probability of falling within a specified period, the probability of not falling within a specified period) output from the output layer 2403 matches (as closely as possible) the actual fall history of each subject (the result of whether or not they have fallen).

[0163] When determining whether or not an intervention by a specialist is necessary based on the prediction result and the diagnosis information, the prediction device 16 determines whether or not an intervention is necessary by using a determination score Y3 AB Calculate the judgment score Y3 AB is the predetermined reference value R3 AB If the score is less than Y3, intervention is deemed unnecessary. AB is the predetermined reference value R3 AB If the condition is more than X3, intervention may be deemed necessary. A is the predicted value, X B is the score value of the diagnostic information, and j and k are coefficients.

[0164] Y3 AB =j×X3 A +k×X B … (5)

[0165] The fourth embodiment also has the same modifications (Modification 1 and Modification 2) as the first to third embodiments, but the description thereof will be omitted.

[0166] An outline of the fifth embodiment (similar to the sixth and seventh embodiments) of the present invention will be described with reference to Fig. 29. Fig. 29 is a conceptual diagram for explaining the outline of the fifth embodiment (similar to the sixth and seventh embodiments).

[0167] As shown in FIG. 29, the system according to the fifth embodiment (similarly to the sixth and seventh embodiments) includes a model generation phase and a prediction phase, similar to the systems according to the first to fourth embodiments. In the model generation phase of the fifth embodiment (similarly to the sixth and seventh embodiments), a trained model (a first trained model described below) that predicts (estimates) diagnostic information and a trained model (a second trained model described below) that predicts the probability of a fall from the diagnostic information (estimated diagnostic information) are generated. In the prediction phase of the fifth embodiment (similarly to the sixth and seventh embodiments), the first trained model is used to predict (estimate) diagnostic information, and the estimated diagnostic information and the second trained model are used to predict the probability of a fall.

[0168] The system according to the fifth embodiment (similar to the sixth and seventh embodiments) predicts (estimates) diagnostic information using at least sensor information obtained by sensing walking. The system according to the fifth embodiment (similar to the sixth and seventh embodiments) also uses at least the estimated diagnostic information to predict the probability of falling within a predetermined period (e.g., within one year) after sensing. Furthermore, the system according to the fifth embodiment (similar to the sixth and seventh embodiments) also uses at least the predicted result of the fall probability (0% to 100%) to determine whether or not professional intervention is required to deal with future falls.

[0169] In addition to sensor information, diagnostic information and questionnaire information may be used to predict diagnostic information, predict the probability of a fall, and determine whether intervention is necessary. In other words, the system according to the fifth embodiment (similar to the sixth and seventh embodiments) may include at least one of the following configurations 10 to 13.

[0170] (Configuration 10) A configuration in which diagnostic information is estimated based on sensor information, a probability of falling is predicted based on the estimated diagnostic information, and whether intervention is necessary is determined based on the predicted result of the probability of falling. (Configuration 11) A configuration in which diagnostic information is estimated based on sensor information, a probability of falling is predicted based on the estimated diagnostic information, and the necessity of intervention is determined based on the predicted result of the probability of falling and questionnaire information. (Configuration 12) A configuration in which diagnostic information is estimated based on sensor information, the probability of a fall is predicted based on the estimated diagnostic information and questionnaire information, and the need for intervention is determined based on the predicted result of the probability of a fall. (Configuration 13) A configuration in which diagnostic information is estimated based on sensor information and questionnaire information, a probability of falling is predicted based on the estimated diagnostic information, and the necessity of intervention is determined based on the predicted result of the probability of falling.

[0171] Machine learning and statistical methods are used to predict diagnostic information. Therefore, prior to predicting diagnostic information, necessary information is collected and a first trained model and a statistical model are created. As an example, in the case of configuration 10, sensor information from each of a large number of subjects and actual diagnostic information from experts for each subject are collected, and a first trained model is created that is trained using feature information based on the sensor information (which may be sensor information) as an input sample and the actual diagnostic information from the experts as an output sample, or a statistical model is created that uses feature information based on the sensor information (which may be sensor information) as an explanatory variable and the actual diagnostic information from the experts as a target variable.

[0172] Machine learning and statistical methods are used to predict the probability of a fall. Therefore, prior to predicting the probability of a fall, necessary information is collected and a second trained model or a statistical model is created. As an example, in the case of configuration 10, actual diagnosis information by an expert for each of a large number of subjects and fall occurrence information (fall history information) indicating whether or not each subject actually fell within a predetermined period (within one year) from the date on which sensor information was acquired are collected, and a second trained model trained using the actual diagnosis information by the expert as an input sample and the fall history information as an output sample, or a statistical model using the actual diagnosis information by the expert as an explanatory variable and the fall history information as a target variable, is created.

[0173] (Fifth embodiment) A fifth embodiment of the present invention will be described below with reference to Figures 30 to 43. The fifth embodiment corresponds to configurations 10 and 11 among configurations 10 to 13 described in Figure 29. Note that the fifth embodiment will be described mainly focusing on the parts that are different from the first to fourth embodiments, and the description of the parts that are common to the first to fourth embodiments will be omitted in part or in whole.

[0174] (Prediction Device 18) The prediction device 18 estimates the subject's diagnostic information using a first trained model 2500 (described later). The prediction device 18 also predicts the subject's fall probability using a second trained model 2600 (described later). The prediction device 18 also determines whether intervention is necessary using at least the predicted result of the fall probability.

[0175] Fig. 30 is a diagram illustrating an example configuration of a prediction device 18. As illustrated in Fig. 30, the prediction device 18 includes an acquisition unit 180, a feature extraction unit 181, a diagnostic value estimation unit (diagnostic value prediction unit) 182, a fall prediction unit (estimation unit) 183, an intervention necessity determination unit 184, an output unit 185, a sensor information storage unit 186, a feature information storage unit 188, a first trained model storage unit 189, an estimation result storage unit 190, a second trained model storage unit 191, a prediction result storage unit 192, and a determination result storage unit 193. Note that, as indicated by the dashed line, the prediction device 18 may further include a questionnaire information storage unit 187.

[0176] The feature extraction unit 181, output unit 185, sensor information storage unit 186, feature information storage unit 188, prediction result storage unit 192, and judgment result storage unit 193 provided in the prediction device 18 of the fifth embodiment are similar to the feature extraction unit 101, output unit 104, sensor information storage unit 105, feature information storage unit 108, prediction result storage unit 110, and judgment result storage unit 111 provided in the prediction device 10 of the first embodiment.

[0177] The acquisition unit 180, like the acquisition unit 100 of the prediction device 10 of the first embodiment, acquires sensor information and stores it in the sensor information storage unit 186. Furthermore, like the acquisition unit 100 of the prediction device 10 of the first embodiment, the acquisition unit 180 may acquire questionnaire information of a questionnaire answered by the subject himself / herself and store it in the questionnaire information storage unit 187. If the acquisition unit 180 does not acquire questionnaire information, the questionnaire information storage unit 187 may not be present.

[0178] The first trained model storage unit 189 stores the first trained model 2500. The first trained model 2500 is generated by a first trained model generation device 38 (described later) and supplied to the prediction device 18.

[0179] FIG. 31 is a diagram illustrating the first trained model 2500. The first trained model 2500 is, for example, a model expressed using a neural network as shown in FIG. 31. As shown in FIG. 31, the first trained model 2500 of the fifth embodiment is a trained model that, when feature information (feature information 1 to feature information N) of a subject is input to the input layer 2501, outputs an estimated result of diagnostic information by an expert for the subject from the output layer 2503. For example, when GARS1 to GARS7 are used as diagnostic information, the first trained model 2500 is a trained model that, when feature information (feature information 1 to feature information N) of the subject is input to the input layer 2501, outputs estimated values ​​of each of GARS1 to GARS7 of the subject from the output layer 2503 as an estimated result of diagnostic information by an expert for the subject.

[0180] The diagnostic value estimation unit 182 estimates (predicts) the subject's diagnostic information (e.g., the values ​​of GARS1 to GARS7) using the feature information stored in the feature information storage unit 188 and the first learned model 2500 stored in the first learned model storage unit 189, and stores the estimation result (estimated diagnostic information, also referred to as an estimated diagnostic value) in the estimation result storage unit 190.

[0181] The second trained model storage unit 191 stores the second trained model 2600. The second trained model 2600 is generated by a second trained model generation device 39 (described later) and supplied to the prediction device 18.

[0182] FIG. 32 is a diagram illustrating the second trained model 2600. The second trained model 2600 is, for example, a model expressed using a neural network as shown in FIG. 32. As shown in FIG. 32, the second trained model 2600 of the fifth embodiment is a trained model that, when an estimated result of diagnostic information of a subject is input to the input layer 2601, outputs from the output layer 2603 the probability that the subject will fall within a predetermined period and the probability that the subject will not fall within a predetermined period. For example, when GARS1 to GARS7 are used as diagnostic information, the second trained model 2600 is a trained model that, when an estimated value of each of GARS1 to GARS7 of the subject is input to the input layer 2601, outputs from the output layer 2603 the probability that the subject will fall within a predetermined period and the probability that the subject will not fall within a predetermined period.

[0183] The fall prediction unit 183 predicts the probability that the subject will fall within a specified period in the future using the estimation results stored in the estimation result memory unit 190 and the second learned model 2600 stored in the second learned model memory unit 191, and stores the prediction result in the prediction result memory unit 192.

[0184] The intervention necessity determination unit 184 determines whether intervention by an expert is necessary based on the prediction result stored in the prediction result storage unit 192. For example, the intervention necessity determination unit 184 may determine that intervention by an expert is unnecessary when the prediction result is less than a predetermined threshold (e.g., 60%), and may determine that intervention by an expert is necessary when the prediction result is equal to or greater than the predetermined threshold. Having determined whether intervention is necessary, the intervention necessity determination unit 184 stores the determination result (necessary / not necessary) in the determination result storage unit 193.

[0185] Furthermore, similar to the prediction device 12 (intervention necessity determination unit 123) of the second embodiment, the intervention necessity determination unit 184 may determine whether or not intervention by an expert is necessary based on the questionnaire information (questionnaire information stored in the questionnaire information storage unit 187) in addition to the prediction result.

[0186] Fig. 33 is a flowchart showing an example of the operation of the prediction device 18. Specifically, the flowchart in Fig. 33 shows an example of the operation of the acquisition unit 180, the feature extraction unit 181, the diagnostic value estimation unit 182, the fall prediction unit 183, the intervention necessity determination unit 184, and the output unit 185.

[0187] The prediction device 18 (acquisition unit 180) acquires sensor information (step S70). The prediction device 18 (acquisition unit 180) stores the sensor information. Subsequently, the prediction device 18 (feature extraction unit 181) extracts features (step S71). The prediction device 18 (feature extraction unit 181) stores the feature information. Subsequently, the prediction device 18 (diagnostic value estimation unit 182) inputs the feature information to the first trained model 2500 to estimate diagnostic information (step S72). Subsequently, the prediction device 18 (diagnostic value estimation unit 182) stores the estimated diagnostic information. The prediction device 18 (fall prediction unit 183) inputs the estimated diagnostic information to the second trained model 2600 to predict the probability of a fall (step S73). The prediction device 18 (fall prediction unit 183) stores the prediction result. Next, prediction device 18 (intervention necessity determination unit 184) determines whether intervention by an expert is necessary (step S74). Prediction device 18 (intervention necessity determination unit 184) stores the determination result. Next, prediction device 18 (output unit 185) outputs the prediction result by the fall prediction unit 183 and the determination result by the intervention necessity determination unit 184 (step S75). Then, this flowchart ends.

[0188] (First trained model generation device 38) The first trained model generation device 38 generates a first trained model 2500 to be used in the prediction device 18. The first trained model generation device 38 trains a learning model using a learning dataset 1500 (described later) derived from multiple measurers (subjects) who acquired sensor information, and generates the first trained model 2500.

[0189] 34 is a diagram illustrating an example configuration of the first trained model generation device 38. As shown in FIG. 34, the first trained model generation device 38 includes an acquisition unit 250, a model generation unit 382, ​​and a storage unit 389.

[0190] The acquiring unit 380 acquires the training dataset 1500 from the outside. After acquiring the training dataset 1500, the acquiring unit 380 stores the training dataset 1500 in the storage unit 2389.

[0191] FIG. 35 is a diagram illustrating a training dataset 1500. The acquiring unit 380 acquires, for example, the training dataset 1500 shown in FIG. 35. The acquiring unit 380 may acquire training samples 1 to N collectively or individually. The acquiring unit 380 may also acquire feature information and diagnostic information individually. In this case, for example, common identification information (such as a subject number) may be added to each of the feature information and diagnostic information so that a single training sample can be generated from the feature information and diagnostic information after acquisition (i.e., so that it can be recognized that they constitute the same training sample (i.e., information on the subject)).

[0192] The training data set 1500 (each training sample) shown in Fig. 35 includes, as an input sample (symbol A), feature information based on each sensor information obtained from each subject, and, as an output sample (symbol B), actual diagnostic information by an expert for each subject.

[0193] Note that feature information corresponding to a portion of the input sample is obtained from sensor information, but the process of obtaining (extracting) the feature information from the sensor information is performed in a device (not shown; a training dataset generation device including a feature extraction unit similar to the feature extraction unit 181 of the prediction device 18 shown in FIG. 30) different from the first trained model generation device 38. In other words, a device (training dataset generation device) different from the first trained model generation device 38 generates the training dataset 1500, and the first trained model generation device 38 acquires the training dataset 1500 from the training dataset generation device.

[0194] The storage unit 389 stores the learning data set 1500 (learning sample 1 to learning sample N in FIG. 35) acquired by the acquisition unit 380. Furthermore, the storage unit 389 stores a learning model and parameters.

[0195] The model generation unit 382 trains a learning model using the learning dataset 1500 stored in the storage unit 389, and generates a first trained model 2500. The model generation unit 382 stores the generated first trained model 2500 in the storage unit 389.

[0196] Specifically, the model generation unit 382 inputs each of the input samples (respective feature information of learning sample 1 to learning sample N in FIG. 35) to the input layer 2501, updates the parameters of the learning model so as to reduce the difference between each output value obtained from the output layer 2503 and each of the output samples (respective diagnostic information of learning sample 1 to learning sample N in FIG. 35), and generates the trained model 2500. For example, the model generation unit 382 may generate the first trained model 2500 using backpropagation, similar to the model generation unit 302 of the model generation device 30 of the first embodiment.

[0197] As described above, the first trained model 2500 is a learning model that has been trained and generated so that when feature information (feature information 1 to feature information N) of each subject is input to the input layer 2501, the estimated value (predicted value) of the diagnostic information (diagnostic information 1 to diagnostic information N) of each subject output from the output layer 2503 matches (matches as closely as possible) the actual diagnostic information (diagnostic information 1 to diagnostic information N) provided by an expert for each subject. For example, when GARS1 to GARS7 are used as diagnostic information, the first trained model 2500 is a learning model that has been trained and generated so that when feature information (feature information 1 to feature information N) of each subject is input to the input layer 2501, the estimated values ​​of GARS1 to GARS7 of each subject output from the output layer 2503 (specifically, for each of GARS1 to GARS7, the probability that the value is 0, the probability that the value is 1, the probability that the value is 2, and the probability that the value is 3) match (as closely as possible) the actual values ​​of GARS1 to GARS7 assigned by experts for each subject.

[0198] Fig. 36 is a flowchart showing an example of operation of the first trained model generation device 38. Specifically, the flowchart in Fig. 36 shows an example of operation of the model generation unit 382. At the start of the flowchart in Fig. 36, it is assumed that the training dataset 1500, training model, and parameters shown in Fig. 35 are stored in the storage unit 389.

[0199] The first trained model generation device 38 (model generation unit 382; the same applies below) inputs each (subject's) input sample (feature information) into the training model and acquires each output value (estimated diagnostic value) from the training model (step S80). The first trained model generation device 38 then calculates the error between each output value (estimated diagnostic value) and each output sample (step S81). The first trained model generation device 38 then determines whether the error is within a predetermined value (step S82). If the error is not within the predetermined value (step S82: NO), the first trained model generation device 38 updates various parameters based on each error (step S83) and returns to step S80. That is, steps S80 to S83 are repeatedly executed until the error is within the predetermined value. If the error is within the predetermined value (step S82: YES), the first trained model generation device 38 ends training (step S84). In other words, the generation of the first trained model 2500 by the first trained model generation device 38 is completed (the trained model in which the various parameters are used as trained parameters is the trained model 2500). Then, this flowchart ends.

[0200] (Modification of the fifth embodiment) The following describes modifications of the fifth embodiment. The modifications may be combined as appropriate.

[0201] (Modification 1 of the fifth embodiment) In the fifth embodiment, it was explained that the process of obtaining feature information from sensor information is performed in a device (learning dataset generation device) different from the first trained model generation device 38, but the first trained model generation device 38 may be responsible for the process of obtaining feature information from sensor information.

[0202] Figure 37 is a diagram illustrating another exemplary configuration of the first trained model generation device 38. As shown in Figure 37, the first trained model generation device 38 may further include an input data storage unit 385 and a feature extraction unit 381 in addition to an acquisition unit 380, a model generation unit 382, ​​and a storage unit 389. The feature extraction unit 381 extracts features from sensor information, similar to the feature extraction unit 181 of the prediction device 18.

[0203] Figure 38 is a diagram illustrating an input dataset 1510. The acquisition unit 380 of the first trained model generation device 38 shown in Figure 34 acquires the training dataset 1500 as shown in Figure 35 and stores it in the storage unit 389, while the acquisition unit 380 of the first trained model generation device 38 shown in Figure 38 acquires the input dataset 1510 as shown in Figure 38 and stores it in the input data storage unit 385.

[0204] The input dataset 1510 (information for each learning sample) shown in Fig. 38 differs from the learning dataset 1500 (each learning sample) shown in Fig. 35 in the following respects: The learning dataset 1500 shown in Fig. 35 includes feature information that is an input sample (symbol A) and diagnostic information that is an output sample (symbol B), whereas the input dataset 1510 shown in Fig. 38 includes sensor information that is information (symbol C) for obtaining the input sample (feature information) and diagnostic information that is an output sample (symbol B).

[0205] The acquiring unit 380 of the first trained model generation device 38 shown in FIG. 37 acquires the input dataset 1510. The acquiring unit 380 may acquire the information for learning sample 1 to learning sample N together or individually. The acquiring unit 380 may also acquire the sensor information and the diagnostic information individually. In this case, for example, common identification information (such as a subject number) may be added to each of the sensor information and the diagnostic information so that information for one learning sample can be generated from the sensor information and the diagnostic information after acquisition (i.e., so that it can be recognized that they constitute information for the same learning sample (i.e., information for one subject)).

[0206] The feature extraction unit 281 extracts features from the sensor information in the input dataset 1510 and generates a training dataset 1500. Thereafter, the first trained model generation device 38 shown in FIG. 37 generates a first trained model 2500, similar to the first trained model generation device 38 shown in FIG. 34.

[0207] (Modification 2 of the fifth embodiment) Fig. 39 is a diagram illustrating another first trained model 2520. The first trained model 2500 shown in Fig. 31 inputs feature information and estimates diagnostic information, but a trained model 2520 as shown in Fig. 39 inputs sensor information and estimates diagnostic information may also be used.

[0208] Figure 40 is a diagram illustrating another training dataset 1520. When using the first trained model 2500 shown in Figure 31, the training dataset 1500 shown in Figure 35 (or the input dataset 1510 shown in Figure 37) is used, but when using the first trained model 2520 shown in Figure 39, the training dataset 1520 shown in Figure 40 is used.

[0209] The training dataset 1520 (respective training samples) shown in FIG. 40 differs from the training dataset 1500 (respective training samples) shown in FIG. 35 in the following respects: The training dataset 1500 shown in FIG. 35 includes feature information, which is an input sample (symbol A), and diagnostic information, which is an output sample (symbol B), whereas the training dataset 1520 shown in FIG. 40 includes sensor information, which is an input sample (symbol A), and diagnostic information, which is an output sample (symbol B). Note that when the first trained model 2520 or the training dataset 1520 is used, feature information is not used, and therefore, needless to say, configurations related to features are not required in each device (prediction device, model generation device).

[0210] (Second trained model generation device 39) The second trained model generation device 39 generates a second trained model 2600 to be used in the prediction device 18. The second trained model generation device 39 trains a learning model using a learning dataset 1600 (described later) derived from a subject, and generates the second trained model 2600. Note that the subject may be a person who has acquired sensor information (measurer), or may be a person who has not acquired sensor information.

[0211] 41 is a diagram illustrating an example configuration of the second trained model generation device 39. As shown in FIG. 41, the second trained model generation device 39 includes an acquisition unit 390, a model generation unit 392, and a storage unit 399.

[0212] The acquiring unit 390 externally acquires the training dataset 1600. After acquiring the training dataset 1600, the acquiring unit 390 stores the training dataset 1600 in the storage unit 399.

[0213] FIG. 42 is a diagram illustrating a training dataset 1600. The acquiring unit 390 acquires, for example, the training dataset 1600 as shown in FIG. 42. The acquiring unit 390 may acquire training samples 1 to N collectively or individually. The acquiring unit 390 may also acquire diagnostic information and performance information individually. In this case, for example, common identification information (such as a subject number) may be added to each of the diagnostic information and performance information so that a single training sample can be generated from the diagnostic information and performance information after acquisition (i.e., so that it can be recognized that they constitute the same training sample (i.e., information on the subject)).

[0214] 42 includes, as an input sample (symbol A), actual diagnostic information by an expert for each subject. Also, as an output sample (symbol B), the training data set 1600 includes actual information (fall history survey information) on whether each subject has fallen within a predetermined period (specifically, within one year from the date of acquisition of each sensor information).

[0215] The storage unit 399 stores the learning data set 1600 (learning sample 1 to learning sample N in FIG. 42) acquired by the acquisition unit 390. Furthermore, the storage unit 399 stores a learning model and parameters.

[0216] The model generation unit 392 trains the learning model using the learning dataset 1600 stored in the storage unit 399, and generates a second trained model 2600. The model generation unit 392 stores the generated second trained model 2600 in the storage unit 399.

[0217] Specifically, the model generation unit 392 inputs each of the input samples (diagnostic information of learning sample 1 to learning sample N in FIG. 42) to the input layer 2601, updates the parameters of the learning model so as to reduce the difference between each of the output values ​​obtained from the output layer 2603 and each of the output samples (performance information of each of learning sample 1 to learning sample N in FIG. 42), and generates the trained model 2600. For example, the model generation unit 392 may generate the trained model 2600 using an error backpropagation algorithm, similar to the model generation unit 302 of the model generation device 30 of the first embodiment.

[0218] As described above, second trained model 2600 is a learning model trained and generated so that when diagnostic information (diagnostic information 1 to diagnostic information N) for each subject is input to input layer 2601, the fall probability for each subject (probability of falling within a predetermined period, probability of not falling within a predetermined period) output from output layer 2603 matches (matches as closely as possible) with each subject's actual fall history (results of whether or not a fall has occurred). For example, if GARS1 to GARS7 are used as diagnostic information, second trained model 2600 is a learning model trained and generated so that when actual diagnostic values ​​of GARS1 to GARS7 for each subject by an expert are input to input layer 2601, the fall probability for each subject output from output layer 2603 matches (matches as closely as possible) with each subject's actual fall history.

[0219] Figure 43 is a flowchart showing an example of operation of the second trained model generation device 39. Specifically, the flowchart in Figure 43 shows an example of operation of the model generation unit 392. It is assumed that at the start of the flowchart in Figure 43, the training dataset 1600, training model, and parameters as shown in Figure 42 are stored in the storage unit 399.

[0220] The second trained model generation device 39 (model generation unit 392; the same applies below) inputs each (each subject's) input sample (diagnosis information) into the training model and acquires each output value from the training model (step S90). Next, the second trained model generation device 39 calculates the error between each output value and each output sample (whether or not a fall occurred) (step S91). Next, the second trained model generation device 39 determines whether the error is within a predetermined value (step S92). If the error is not within the predetermined value (step S92: NO), the second trained model generation device 39 updates various parameters based on each error (step S93) and returns to step S90. That is, steps S90 to S93 are repeatedly executed until the error is within the predetermined value. If the error is within the predetermined value (step S92: YES), the second trained model generation device 39 ends training (step S94). In other words, the generation of the second trained model 2600 by the second trained model generation device 39 is completed (the trained model in which the various parameters are trained parameters is the trained model 2600). Then, this flowchart ends.

[0221] (Sixth embodiment) Next, a sixth embodiment of the present invention will be described with reference to Figures 44 to 49. The sixth embodiment corresponds to configuration 12 among configurations 10 to 13 described in Figure 29. Note that the sixth embodiment will be described mainly focusing on the parts that are different from the fifth embodiment, and the description of the parts that are common to the fifth embodiment will be omitted in part or in whole.

[0222] (Prediction device 20) The prediction device 20 estimates diagnostic information of the subject using a first trained model 2500 (FIG. 31). The prediction device 20 also predicts the subject's probability of falling using a second trained model 2800 (described later). The prediction device 20 also determines whether intervention is necessary using the predicted result of the fall probability.

[0223] Fig. 44 is a diagram illustrating a configuration example of prediction device 20. As illustrated in Fig. 44, prediction device 20 includes an acquisition unit 200, a feature extraction unit 201, a diagnostic value estimation unit (diagnostic value prediction unit) 202, a fall prediction unit (estimation unit) 203, an intervention necessity determination unit 204, an output unit 205, a sensor information storage unit 206, a questionnaire information storage unit 207, a feature information storage unit 208, a first trained model storage unit 209, an estimation result storage unit 210, a second trained model storage unit 211, a prediction result storage unit 212, and a determination result storage unit 213.

[0224] The feature extraction unit 201, diagnostic value estimation unit 202, output unit 205, sensor information storage unit 206, questionnaire information storage unit 207, feature information storage unit 208, first learned model storage unit 209, estimation result storage unit 210, prediction result storage unit 212, and judgment result storage unit 213 provided in the prediction device 20 of the sixth embodiment are similar to the feature extraction unit 181, diagnostic value estimation unit 182, output unit 185, sensor information storage unit 186, questionnaire information storage unit 187, feature information storage unit 188, first learned model storage unit 189, estimation result storage unit 190, prediction result storage unit 192, and judgment result storage unit 193 provided in the prediction device 18 of the fifth embodiment.

[0225] The acquiring unit 200, like the acquiring unit 180 of the prediction device 18 of the fifth embodiment, acquires sensor information and stores it in the sensor information storage unit 206. In addition, the acquiring unit 200 acquires questionnaire information of a questionnaire answered by the subject himself / herself and stores it in the questionnaire information storage unit 207.

[0226] The second trained model storage unit 211 stores the second trained model 2800. The second trained model 2800 is generated by a second trained model generation device 41 (described later) and supplied to the prediction device 20.

[0227] FIG. 45 is a diagram illustrating the second trained model 2800. The second trained model 2800 is, for example, a model expressed using a neural network as shown in FIG. 45. As shown in FIG. 45, the second trained model 2800 of the sixth embodiment is a trained model that, when the estimation results of the subject's diagnostic information and questionnaire information (questionnaire information 1 to questionnaire information N) are input to the input layer 2801, outputs from the output layer 2803 the probability that the subject will fall within a predetermined period and the probability that the subject will not fall within a predetermined period. For example, when GARS1 to GARS7 are used as the diagnostic information, the second trained model 2800 is a trained model that, when the estimation values ​​of GARS1 to GARS7 of the subject and questionnaire information (questionnaire information 1 to questionnaire information N) are input to the input layer 2801, outputs from the output layer 2803 the probability that the subject will fall within a predetermined period and the probability that the subject will not fall within a predetermined period.

[0228] The fall prediction unit 203 predicts the probability that the subject will fall within a specified period in the future using the estimation results stored in the estimation result memory unit 210, the questionnaire information stored in the questionnaire information memory unit 207, and the second trained model 2800 stored in the second trained model memory unit 211, and stores the prediction result in the prediction result memory unit 212.

[0229] The intervention necessity determination unit 204 determines whether intervention by an expert is necessary based on the prediction result stored in the prediction result storage unit 212. For example, the intervention necessity determination unit 204 may determine that intervention by an expert is unnecessary when the prediction result is less than a predetermined threshold (e.g., 60%), and may determine that intervention by an expert is necessary when the prediction result is equal to or greater than the predetermined threshold. Having determined whether intervention is necessary, the intervention necessity determination unit 204 stores the determination result (necessary / not necessary) in the determination result storage unit 213.

[0230] Fig. 46 is a flowchart showing an example of the operation of the prediction device 20. Specifically, the flowchart in Fig. 46 shows an example of the operation of the acquisition unit 200, the feature extraction unit 201, the diagnostic value estimation unit 202, the fall prediction unit 203, the intervention necessity determination unit 204, and the output unit 205.

[0231] The prediction device 20 (acquisition unit 200) acquires sensor information and questionnaire information (step S100). The prediction device 14 (acquisition unit 200) stores the sensor information and questionnaire information. Subsequently, the prediction device 20 (feature extraction unit 201) extracts features (step S101). The prediction device 20 (feature extraction unit 201) stores the feature information. Subsequently, the prediction device 20 (diagnostic value estimation unit 202) inputs the feature information to the first trained model 2500 to estimate diagnostic information (step S102). Subsequently, the prediction device 20 (diagnostic value estimation unit 202) stores the estimated diagnostic information. Subsequently, the prediction device 20 (fall prediction unit 203) inputs the estimated diagnostic information and the questionnaire information to the second trained model 2800 to predict the probability of a fall (step S103). The prediction device 20 (fall prediction unit 203) stores the prediction result. Next, the prediction device 20 (intervention necessity determination unit 204) determines whether intervention by a specialist is necessary (step S104). The prediction device 20 (intervention necessity determination unit 204) stores the determination result. Next, the prediction device 20 (output unit 205) outputs the prediction result by the fall prediction unit 203 and the determination result by the intervention necessity determination unit 204 (step S105). Then, this flowchart ends.

[0232] (Second trained model generation device 41) The second trained model generation device 41 generates a second trained model 2800 to be used in the prediction device 20. The second trained model generation device 41 trains a learning model using a learning dataset 1800 (described later) derived from a subject, and generates the second trained model 2800. Note that the subject may be a person who has acquired sensor information (measurer), or may be a person who has not acquired sensor information.

[0233] 47 is a diagram illustrating an example configuration of the second trained model generation device 41. As shown in FIG. 47, the second trained model generation device 41 includes an acquisition unit 410, a model generation unit 412, and a storage unit 419.

[0234] The acquisition unit 410 externally acquires the training dataset 1800. After acquiring the training dataset 1800, the acquisition unit 410 stores the training dataset 1800 in the storage unit 419.

[0235] FIG. 48 is a diagram illustrating a training dataset 1800. The acquiring unit 410 acquires, for example, the training dataset 1800 as shown in FIG. 48. The acquiring unit 410 may acquire training samples 1 to N collectively or individually. The acquiring unit 410 may also acquire the diagnostic information, questionnaire information, and performance information individually. In this case, for example, common identification information (such as a subject number) may be added to each of the diagnostic information, questionnaire information, and performance information so that one training sample can be generated from the diagnostic information, questionnaire information, and performance information after acquisition (i.e., so that it can be recognized that they constitute the same training sample (i.e., information on the subject)).

[0236] 48 includes, as an input sample (symbol A), actual diagnostic information by an expert and questionnaire information for each subject. Also, as an output sample (symbol B), it includes actual information (fall history survey information) on whether each subject has fallen within a predetermined period (specifically, within one year from the date on which each sensor information was acquired).

[0237] The storage unit 419 stores the learning data set 1800 (learning sample 1 to learning sample N in FIG. 48) acquired by the acquisition unit 410. Furthermore, the storage unit 419 stores a learning model and parameters.

[0238] The model generation unit 412 trains a learning model using the learning dataset 1800 stored in the storage unit 419, and generates a trained model 2800. The model generation unit 412 stores the generated trained model 2800 in the storage unit 419.

[0239] Specifically, the model generation unit 412 inputs each of the input samples (diagnosis information and questionnaire information of learning sample 1 to learning sample N in FIG. 48) to the input layer 2801, updates the parameters of the learning model so as to reduce the difference between each output value obtained from the output layer 2803 and each of the output samples (performance information of each of learning sample 1 to learning sample N in FIG. 48), and generates the trained model 2800. For example, the model generation unit 412 may generate the trained model 2800 using an error backpropagation algorithm, similar to the model generation unit 302 of the model generation device 30 of the first embodiment.

[0240] As described above, second trained model 2800 is a learning model trained and generated so that the fall probability of each subject (probability of falling within a predetermined period, probability of not falling within a predetermined period) output from output layer 2803 when diagnostic information (diagnostic information 1 to diagnostic information N) and questionnaire information (questionnaire information 1 to questionnaire information N) of each subject are input to input layer 2801 matches (matches as closely as possible) with the actual fall history of each subject (results of whether or not a fall occurred). For example, if GARS1 to GARS7 are used as diagnostic information, second trained model 2800 is a learning model trained and generated so that the fall probability of each subject output from output layer 2803 when actual diagnostic values ​​of GARS1 to GARS7 by experts for each subject and questionnaire information (questionnaire information 1 to questionnaire information N) are input to input layer 2801 matches (matches as closely as possible) with the actual fall history of each subject.

[0241] Figure 49 is a flowchart showing an example of operation of the second trained model generation device 41. Specifically, the flowchart in Figure 49 shows an example of operation of the model generation unit 412. It is assumed that at the start of the flowchart in Figure 49, the training dataset 1800, training model, and parameters as shown in Figure 48 are stored in the storage unit 419.

[0242] The second trained model generation device 41 (model generation unit 412; the same applies below) inputs input samples (diagnosis information, questionnaire information) of each subject (each subject) into the learning model and acquires each output value (estimated diagnosis value) from the learning model (step S110). Next, the second trained model generation device 41 calculates the error between each output value (estimated diagnosis value) and each output sample (presence or absence of fall) (step S111). Next, the second trained model generation device 41 determines whether the error is within a predetermined value (step S112). If the error is not within the predetermined value (step S112: NO), the second trained model generation device 41 updates various parameters based on each error (step S113) and returns to step S110. In other words, steps S110 to S113 are repeatedly executed until the error is within the predetermined value. If the error is within the predetermined value (step S112: YES), the second trained model generation device 41 ends training (step S114). That is, the generation of the second trained model 2800 by the second trained model generation device 41 is completed (the trained model with the various parameters as trained parameters is the trained model 2800). Then, this flowchart ends.

[0243] The sixth embodiment also has the same modifications as the fifth embodiment (Modification 1 and Modification 2), but the description thereof will be omitted.

[0244] (Seventh embodiment) Next, a seventh embodiment of the present invention will be described with reference to Figures 50 to 55. The seventh embodiment corresponds to configuration 13 among configurations 10 to 13 described in Figure 29. Note that the seventh embodiment will be described mainly focusing on the parts that are different from the fifth embodiment, and the description of the parts that are common to the fifth embodiment will be omitted in part or in whole.

[0245] (Prediction Device 22) The prediction device 22 estimates the subject's diagnostic information using a first trained model 2900 (described later). The prediction device 22 also predicts the subject's fall probability using a second trained model 2600 (FIG. 32). The prediction device 22 also determines whether intervention is necessary using the predicted fall probability.

[0246] 50 is a diagram illustrating a configuration example of the prediction device 22. As illustrated in FIG. 50, the prediction device 22 includes an acquisition unit 220, a feature extraction unit 221, a diagnostic value estimation unit (diagnostic value prediction unit) 222, a fall prediction unit (estimation unit) 223, an intervention necessity determination unit 224, an output unit 225, a sensor information storage unit 226, a questionnaire information storage unit 227, a feature information storage unit 228, a first trained model storage unit 229, an estimation result storage unit 230, a second trained model storage unit 231, a prediction result storage unit 232, and a determination result storage unit 233.

[0247] The feature extraction unit 221, fall prediction unit 223, output unit 225, sensor information storage unit 226, questionnaire information storage unit 227, feature information storage unit 228, estimation result storage unit 230, second learned model storage unit 231, prediction result storage unit 232, and judgment result storage unit 233 provided in the prediction device 22 of the seventh embodiment are similar to the feature extraction unit 181, fall prediction unit 183, output unit 185, sensor information storage unit 186, questionnaire information storage unit 187, feature information storage unit 188, estimation result storage unit 190, second learned model storage unit 191, prediction result storage unit 192, and judgment result storage unit 193 provided in the prediction device 18 of the fifth embodiment.

[0248] The acquisition unit 220, like the acquisition unit 180 of the prediction device 18 of the fifth embodiment, acquires sensor information obtained from the sensor shoes and stores it in the sensor information storage unit 226. In addition, the acquisition unit 220 acquires questionnaire information of a questionnaire answered by the subject himself / herself and stores it in the questionnaire information storage unit 227.

[0249] The first trained model storage unit 229 stores the first trained model 2900. The first trained model 2900 is generated by a first trained model generation device 42 (described later) and supplied to the prediction device 22.

[0250] FIG. 51 is a diagram illustrating a first trained model 2900. The first trained model 2900 is, for example, a model expressed using a neural network as shown in FIG. 51. As shown in FIG. 51, the first trained model 2900 of the seventh embodiment is a trained model that, when feature information (feature information 1 to feature information N) of a subject and questionnaire information (questionnaire information 1 to questionnaire information N) are input to the input layer 2901, outputs an estimated result of diagnostic information by an expert for the subject from the output layer 2903. For example, when GARS1 to GARS7 are used as diagnostic information, the first trained model 2900 is a trained model that, when feature information (feature information 1 to feature information N) of the subject and questionnaire information (questionnaire information 1 to questionnaire information N) are input to the input layer 2901, outputs estimated values ​​of each of GARS1 to GARS7 of the subject from the output layer 2903 as an estimated result of diagnostic information by an expert for the subject.

[0251] The diagnostic value estimation unit 222 estimates (predicts) the subject's diagnostic information (e.g., the values ​​of GARS1 to GARS7) using the feature information stored in the feature information storage unit 228 and the first trained model 2900 stored in the first trained model storage unit 229, and stores the estimation result in the estimation result storage unit 230.

[0252] The intervention necessity determination unit 224 determines whether intervention by an expert is necessary based on the prediction result stored in the prediction result storage unit 232. For example, the intervention necessity determination unit 224 may determine that intervention by an expert is unnecessary when the prediction result is less than a predetermined threshold (e.g., 60%), and may determine that intervention by an expert is necessary when the prediction result is equal to or greater than the predetermined threshold. Having determined whether intervention is necessary, the intervention necessity determination unit 224 stores the determination result (necessary / not necessary) in the determination result storage unit 233.

[0253] Fig. 52 is a flowchart showing an example of the operation of the prediction device 22. Specifically, the flowchart in Fig. 52 shows an example of the operation of the acquisition unit 220, the feature extraction unit 221, the diagnostic value estimation unit 222, the fall prediction unit 223, the intervention necessity determination unit 224, and the output unit 225.

[0254] The prediction device 22 (acquisition unit 220) acquires sensor information and questionnaire information (step S120). The prediction device 22 (acquisition unit 220) stores the sensor information and questionnaire information. Subsequently, the prediction device 22 (feature extraction unit 221) extracts features (step S121). The prediction device 22 (feature extraction unit 221) stores the feature information. Subsequently, the prediction device 22 (diagnostic value estimation unit 222) inputs the feature information and questionnaire information to the first trained model 2900 to estimate diagnostic information (step S122). Subsequently, the prediction device 22 (diagnostic value estimation unit 222) stores the estimated diagnostic information. The prediction device 22 (fall prediction unit 223) inputs the estimated diagnostic information to the second trained model 2600 to predict the probability of a fall (step S123). The prediction device 22 (fall prediction unit 223) stores the prediction result. Next, the prediction device 22 (intervention necessity determination unit 224) determines whether intervention by a specialist is necessary (step S124). The prediction device 22 (intervention necessity determination unit 224) stores the determination result. Next, the prediction device 22 (output unit 225) outputs the prediction result by the fall prediction unit 223 and the determination result by the intervention necessity determination unit 224 (step S125). Then, this flowchart ends.

[0255] (First trained model generation device 42) The first trained model generation device 42 generates a first trained model 2900 to be used in the prediction device 22. The first trained model generation device 42 trains a learning model using a learning dataset 1900 (described later) derived from multiple measurers (subjects) who acquired sensor information, and generates the first trained model 2900.

[0256] 53 is a diagram illustrating an example configuration of the first trained model generation device 42. As shown in FIG. 53, the first trained model generation device 42 includes an acquisition unit 420, a model generation unit 422, and a storage unit 429.

[0257] The acquisition unit 420 externally acquires the training dataset 1900. After acquiring the training dataset 1900, the acquisition unit 420 stores the training dataset 1900 in the storage unit 429.

[0258] FIG. 54 is a diagram illustrating a training dataset 1900. The acquiring unit 420 acquires, for example, the training dataset 1900 as shown in FIG. 54. The acquiring unit 420 may acquire training samples 1 to N collectively or individually. The acquiring unit 420 may also acquire feature information, questionnaire information, and diagnostic information individually. In this case, for example, common identification information (such as a subject number) may be added to each of the feature information, questionnaire information, and diagnostic information so that one training sample can be generated from the feature information, questionnaire information, and diagnostic information after acquisition (that is, so that it can be recognized that they constitute the same training sample (that is, information on the subject)).

[0259] The learning data set 1500 (each learning sample) shown in Fig. 54 includes, as an input sample (symbol A), feature information based on each sensor information obtained from each subject and questionnaire information obtained from each subject, and, as an output sample (symbol B), actual diagnostic information by an expert for each subject.

[0260] Note that feature information corresponding to a portion of the input sample is obtained from sensor information, but the process of obtaining (extracting) the feature information from the sensor information is performed in a device (not shown; a training dataset generation device having a feature extraction unit similar to the feature extraction unit 221 of the prediction device 22 shown in FIG. 50) different from the first trained model generation device 42. In other words, a device (training dataset generation device) different from the first trained model generation device 42 generates the training dataset 1900, and the first trained model generation device 42 acquires the training dataset 1900 from the training dataset generation device.

[0261] The storage unit 429 stores the learning data set 1900 (learning sample 1 to learning sample N in FIG. 54) acquired by the acquisition unit 420. Furthermore, the storage unit 429 stores a learning model and parameters.

[0262] The model generation unit 422 trains a learning model using the learning dataset 1900 stored in the storage unit 429, and generates a first trained model 2900. The model generation unit 422 stores the generated first trained model 2900 in the storage unit 429.

[0263] Specifically, the model generation unit 422 inputs each of the input samples (feature information of each of learning samples 1 to N in FIG. 54 and questionnaire information) to the input layer 2901, updates the parameters of the learning model so as to reduce the difference between each of the output values ​​obtained from the output layer 2903 and each of the output samples (diagnosis information of each of learning samples 1 to N in FIG. 54), and generates the trained model 2900. For example, the model generation unit 422 may generate the first trained model 2900 using an error backpropagation algorithm, similar to the model generation unit 302 of the model generation device 30 of the first embodiment.

[0264] As described above, the first trained model 2900 is a learning model that has been trained and generated so that when the feature information (feature information 1 to feature information N) and questionnaire information (questionnaire information 1 to questionnaire information N) of each subject are input to the input layer 2901, the estimated values ​​(predicted values) of the diagnostic information (diagnostic information 1 to diagnostic information N) of each subject output from the output layer 2903 match (match as closely as possible) the actual diagnostic information (diagnostic information 1 to diagnostic information N) provided by an expert for each subject. For example, when GARS1 to GARS7 are used as diagnostic information, the first trained model 2900 is a learning model that has been trained and generated so that when feature information (feature information 1 to feature information N) and questionnaire information (questionnaire information 1 to questionnaire information N) of each subject are input to the input layer 2901, the estimated values ​​of GARS1 to GARS7 of each subject output from the output layer 2903 (specifically, for each of GARS1 to GARS7, the probability that the value is 0, the probability that the value is 1, the probability that the value is 2, and the probability that the value is 3) match (as closely as possible) the actual values ​​of GARS1 to GARS7 assigned by experts for each subject.

[0265] Figure 55 is a flowchart showing an example of operation of the first trained model generation device 42. Specifically, the flowchart in Figure 55 shows an example of operation of the model generation unit 422. It is assumed that at the start of the flowchart in Figure 55, the training dataset 1900, training model, and parameters as shown in Figure 54 are stored in the storage unit 429.

[0266] The first trained model generation device 42 (model generation unit 422; the same applies below) inputs input samples (feature information, questionnaire information) of each subject (each subject) into the training model and acquires each output value (estimated diagnostic value) from the training model (step S130). Next, the first trained model generation device 42 calculates the error between each output value (estimated diagnostic value) and each output sample (step S131). Next, the first trained model generation device 42 determines whether the error is within a predetermined value (step S132). If the error is not within the predetermined value (step S132: NO), the first trained model generation device 42 updates various parameters based on each error (step S133) and returns to step S130. That is, steps S130 to S133 are repeatedly executed until the error is within the predetermined value. If the error is within the predetermined value (step S132: YES), the first trained model generation device 42 ends training (step S134). That is, the generation of the first trained model 2900 by the first trained model generation device 42 is completed (the trained model with the various parameters as trained parameters is the trained model 2900). Then, this flowchart ends.

[0267] The seventh embodiment also has the same modifications (Modification 1 and Modification 2) as the fifth embodiment, but the description thereof will be omitted.

[0268] (Other methods for predicting the probability of falling) In the above embodiments (first to seventh embodiments), examples have been described in which a neural network is used as machine learning to predict the probability of falling, but the machine learning method is not limited to a neural network. For example, a support vector machine (linear kernel, non-linear kernel), a random forest, or gradient boosting may be used, or a majority model (stacking) that combines these may be used.

[0269] Furthermore, the probability of falling may be predicted using a statistical model (statistical analysis, mathematical formula). For example, the probability of falling may be predicted by logistic regression analysis. For example, in the first embodiment, a logistic function may be created (estimated) with feature amount information as an explanatory variable and the presence or absence of a fall as an objective variable. In the second embodiment, a logistic function may be created with feature amount information and diagnostic information as explanatory variables and the presence or absence of a fall as an objective variable. In the third embodiment, a logistic function may be created with feature amount information, diagnostic information, and questionnaire information as explanatory variables and the presence or absence of a fall as an objective variable. The same applies to the other embodiments. The statistical model may be stored in a model storage unit (model storage unit 109, 129, 149, etc.).

[0270] Which of the multiple explanatory variable candidates to use may be determined according to AIC (Akaike Information Criterion) (the combination with the smallest AIC is selected). That is, in the first embodiment, which of feature amount information 1 to feature amount information N to use may be determined according to AIC, in the second embodiment, which of feature amount information 1 to feature amount information N and diagnostic information 1 to diagnostic information N to use may be determined according to AIC, and in the third embodiment, which of feature amount information 1 to feature amount information N, diagnostic information 1 to diagnostic information N and questionnaire information 1 to questionnaire information N to use may be determined according to AIC. The same applies to the other embodiments. The statistical model may be stored in a model storage unit (model storage unit 109, 129, 149, etc.).

[0271] Furthermore, in each embodiment, sensor information may be used as an explanatory variable instead of feature amount information. That is, in the first embodiment, a logistic function may be created in which sensor information is used as an explanatory variable and whether or not a fall has occurred is used as an objective variable. In the second embodiment, a logistic function may be created in which sensor information and diagnostic information are used as explanatory variables and whether or not a fall has occurred is used as an objective variable. In the third embodiment, a logistic function may be created in which sensor information, diagnostic information, and questionnaire information are used as explanatory variables and whether or not a fall has occurred is used as an objective variable. The same applies to the other embodiments.

[0272] Fig. 56 is an explanatory diagram illustrating model evaluation (evaluation by parameters used). The upper part of Fig. 56 shows AIC values ​​when feature amount information based only on acceleration sensors is used as feature amount information, and when feature amount information based on all sensors (acceleration sensors, pressure sensors, bending sensors) is used, in each of the first embodiment (a case in which the probability of falling is predicted based on feature amount information), the second embodiment (a case in which the probability of falling is predicted based on feature amount information + diagnostic information), and the third embodiment (a case in which the probability of falling is predicted based on feature amount information + diagnostic information + questionnaire).

[0273] The graph on the lower left of Figure 56 is a plot of predicted values ​​and observed values ​​(actual results) in the first embodiment. The graph in the lower center of Figure 56 is a plot of predicted values ​​and observed values ​​in the second embodiment. The graph on the lower right of Figure 56 is a plot of predicted values ​​and observed values ​​in the third embodiment. For convenience of the drawing, this is shown in another figure, Figure 57(H), but ideally (high accuracy) the plots should be concentrated in the upper right and lower left (symbol A in Figure 57(H)) and not plotted in the lower right and upper left (symbol A in Figure 56(H)).

[0274] According to Fig. 56, the accuracy of prediction of the probability of falling increases in the order of the first embodiment, the second embodiment, and the third embodiment. In other words, prediction using feature amount information and diagnostic information is more accurate than prediction using feature amount information alone, and further, prediction using feature amount information, diagnostic information, and questionnaire information is more accurate than prediction using feature amount information and diagnostic information. Also, according to Fig. 56, in each embodiment, the AIC value is smaller when feature amount information based on all sensors is used than when feature amount information based on an acceleration sensor is used, which is preferable.

[0275] FIG. 57 is an explanatory diagram illustrating model evaluation (evaluation by model used). FIG. 57(A) is a plot of predicted values ​​and observed values ​​using a model using logistic regression (logistic function). FIG. 57(B) is a plot of predicted values ​​and observed values ​​using a model using a support vector machine (linear kernel). FIG. 57(C) is a plot of predicted values ​​and observed values ​​using a model using a support vector machine (nonlinear kernel). FIG. 57(D) is a plot of predicted values ​​and observed values ​​using a model using random forest. FIG. 57(E) is a plot of predicted values ​​and observed values ​​using a model using a neural network. FIG. 57(F) is a plot of predicted values ​​and observed values ​​using a model using gradient boosting. FIG. 57(G) is a plot of predicted values ​​and observed values ​​using a model using stacking. FIG. 57(H) is an explanatory diagram illustrating how to read the diagram. As mentioned above, the ideal situation (high accuracy) would be for the plots to be concentrated in the upper right and lower left (symbol A in Figure 57(H)) and not plotted in the lower right and upper left (symbol A in Figure 57(H)). The case using a neural network comes closest to this, resulting in high prediction accuracy.

[0276] The probability of falling may be predicted using a simple linear model. For example, in the first embodiment, a linear equation with each piece of feature amount information (feature amount information 1, feature amount information 2, ...) as a variable may be used as a formula for calculating the probability of falling (or may be approximated to a linear equation).

[0277] The probability of a fall may also be predicted using a conversion table. For example, in the first embodiment, the probability of a fall may be predicted using a conversion table that associates a combination of feature amount information values ​​(value ranges) with the probability of a fall (for example, a conversion table that stores a fall probability of △△% corresponding to a value of feature amount information 1, XX, ..., a value of feature amount information 2, XX, ..., a fall probability of □□% corresponding to a value of feature amount information 1, XX, ..., a value of feature amount information 2, XX, ..., etc.). The same applies to the other embodiments. Note that when a linear model or a conversion table is used, the number of pieces of feature amount information may be reduced compared to when machine learning or a nonlinear model is used. The conversion table may be stored in a model storage unit (such as the model storage unit 109).

[0278] (Other methods for determining whether intervention is necessary) In the first embodiment, an example was described in which, when determining whether intervention is necessary based on other information (diagnostic information, questionnaire information) in addition to the prediction results of fall prediction, the information is scored to determine whether intervention is necessary, but the method of determining whether intervention is necessary based on other information in addition to the prediction results of fall prediction is not limited to this.

[0279] For example, a statistical model may be used to determine whether intervention is necessary. As an example, when determining whether intervention is necessary based on the prediction result and diagnostic information, instead of scoring the prediction result and diagnostic information to determine whether intervention is necessary, a logistic function may be created using the prediction result and diagnostic information as explanatory variables and the need for intervention as the objective variable to determine whether intervention is necessary. Note that which of multiple explanatory variable candidates (N explanatory variables of feature amount information 1 to feature amount information N excluding one of the prediction results) to use may be determined according to AIC. The same applies to the case where the need for intervention is determined based on other information (questionnaire information) in addition to the prediction result of fall prediction in the second embodiment, or the case where the need for intervention is determined based on other information (diagnostic information) in addition to the prediction result of fall prediction in the fourth embodiment.

[0280] Furthermore, the need for intervention may be determined using a machine learning technique such as a neural network. As an example, the need for intervention may be determined using a trained model that outputs a determination result on the need for intervention for a subject from an output layer when sensor information (or feature information) of the subject is input to an input layer. Note that, after acquiring feature information (or sensor information) for multiple subjects, intervention need information (e.g., information evaluated based on multiple diagnostic information, multiple questionnaire information, etc.) evaluated by an evaluator with reference to the walking behavior within a specific period may be acquired, and a training dataset may be generated that includes the detection information of each subject as an input sample and the intervention need information of each subject as an output sample, and the trained model may be generated using the training dataset.

[0281] As described above, Embodiments 1 to 7 have been described, but any of the embodiments can suitably predict future falls, etc. Furthermore, Embodiments 5 to 7 are configured to estimate (predict) diagnostic information from an expert and use the estimated diagnostic information to predict future falls, etc., and are characterized in that no expert intervention is required when predicting future falls, etc. (i.e., in the prediction phase).

[0282] Furthermore, although the configurations of each of the first to seventh embodiments have been described, the specific configurations are not limited to these embodiments, and include designs within the scope of the gist of the present invention. For example, the following configurations (1) to (16) are also possible. Note that the following (1) to (16) may be combined as appropriate.

[0283] (1) In the above embodiment, an example was described in which sensor shoes were used to sense the legs and use the sensed data to predict the probability of falling. However, as mentioned above, other body parts may be sensed and used to predict the probability of falling. For example, sensors may be worn on the arms or head instead of or in addition to the legs, and the sensor information (feature information based on the sensor information) may be used to predict the probability of falling.

[0284] (2) In the above embodiment, GARS is used as the diagnostic information, but GARS is only an example, and the diagnostic information does not have to be GARS. For example, other diagnostic criteria may be used, or original diagnostic items may be set.

[0285] (3) In the above embodiments (for example, the first to seventh embodiments), the determination of whether intervention is necessary is a binary determination of necessary / unnecessary, but the determination may be three or more values. For example, a four-value determination (for example, determination A: no problem, determination B: follow-up observation required, determination C: consult a hospital or the like, determination D: visit a hospital immediately) may be made. In the case of the above four-value determination, it is sufficient to set three thresholds or determination reference values ​​indicating the boundary between determination A and determination B, the boundary between determination B and determination C, and the boundary between determination C and determination D.

[0286] (4) In the above embodiments (for example, the first to seventh embodiments), the need for intervention is determined, but the need for intervention does not have to be determined. That is, in the prediction phase, the prediction device 10 (as well as the prediction device 12, prediction device 14, prediction device 16, prediction device 18, prediction device 20, and prediction device 22) does not have to determine the need for intervention.

[0287] (5) In the above embodiments (for example, the first to fourth embodiments), examples have been described in which an expert checks captured images (images capturing the walking behavior) and creates diagnostic information, but the expert may also directly check the walking behavior and create diagnostic information.

[0288] (6) In the above embodiments (e.g., first to seventh embodiments), examples have been described in which an expert checks captured images to create diagnostic information, or in which diagnostic information is estimated using machine learning techniques, but the diagnostic information may be generated (estimated) by analyzing captured images. Furthermore, the diagnostic information may be generated (estimated) using machine learning (e.g., CNN (Convolutional Neural Network)) and captured images.

[0289] (7) In the above embodiments (e.g., the first to seventh embodiments), examples have been described in which a trained model, a statistical model, etc. are generated based on sensing a single walking event. However, a trained model, a statistical model, etc. may be generated based on sensing multiple walking events (multiple events within one year of the initial sensing). Note that the input sample of the training dataset for generating the trained model includes multiple sensor information (or feature information). Furthermore, the sensor information (or feature information) used to predict the probability of falling is multiple sensor information (or feature information).

[0290] (8) In the above embodiments (e.g., the second and third embodiments), examples have been described in which a trained model, a statistical model, etc. are generated based on a single expert diagnosis. However, a trained model, a statistical model, etc. may be generated based on multiple expert diagnoses (multiple diagnoses within one year of the first diagnosis). Note that the input samples of the training dataset for generating the trained model include multiple diagnostic information. Furthermore, the diagnostic information used to predict the probability of a fall is diagnostic information from multiple times.

[0291] (9) In the above embodiments (e.g., the third, fourth, sixth, and seventh embodiments), examples have been described in which a trained model, a statistical model, etc. is generated based on a single questionnaire. However, a trained model, a statistical model, etc. may be generated based on multiple questionnaires (multiple questionnaires within one year of the first questionnaire). Note that the input sample of the training dataset for generating the trained model includes information from multiple questionnaires. Furthermore, the questionnaire information used to predict the probability of a fall is information from multiple questionnaires.

[0292] (10) In the prediction devices described in the above embodiments (e.g., the first to seventh embodiments), the function of predicting the probability of a fall and the function of determining whether intervention is necessary are performed by separate components (a fall prediction unit and an intervention necessity determination unit), but these functions may be performed by a single component. For example, in the first embodiment, the fall prediction unit 102 of the prediction device 10 may predict the probability of a fall and determine whether intervention is necessary. Note that in addition to the prediction of the probability of a fall (prediction result), the determination of whether intervention is necessary (determination result) may also be considered as prediction information regarding future falls.

[0293] (11) In the prediction devices described in the above embodiments (e.g., the fifth to seventh embodiments), the function of estimating diagnostic information and the function of predicting the probability of a fall are performed by separate components (diagnostic value estimation unit, fall prediction unit), but these functions may be performed by a single component. For example, in the fifth embodiment, the diagnostic value estimation unit 182 of the prediction device 18 may estimate the diagnostic information and predict the probability of a fall. Also, a single component may be responsible for the function of estimating diagnostic information, the function of predicting the probability of a fall, and the function of determining whether intervention is necessary. Note that in addition to the prediction of the probability of a fall (prediction result), the estimation of diagnostic information (estimation result) may also be considered as predictive information regarding future falls.

[0294] (12) The functions of each device described in the above embodiments (e.g., the first to seventh embodiments) may be distributed among multiple devices. For example, the first embodiment uses the prediction device 10 equipped with a function for predicting the probability of a fall (the fall prediction unit 102) and a function for determining whether intervention is necessary (the intervention necessity determination unit 103). However, a fall probability prediction device having a function for predicting the probability of a fall and an intervention necessity determination device having a function for determining whether intervention is necessary may be used. Furthermore, for example, the fifth embodiment uses the prediction device 18 equipped with a function for estimating diagnostic information (the diagnostic value estimation unit 182), a function for predicting the probability of a fall (the fall prediction unit 183), and a function for determining whether intervention is necessary (the intervention necessity determination unit 184). However, a diagnostic information estimation device having a function for estimating diagnostic information, a fall probability prediction device having a function for predicting the probability of a fall, and an intervention necessity determination device having a function for determining whether intervention is necessary may be used.

[0295] (13) The functions of each device described in the above embodiments (e.g., the first to seventh embodiments) may be concentrated in a single device. For example, in the first embodiment, a model generation device 30 and a prediction device 10 are used, but a device combining the functions of the model generation device 30 and the prediction device 10 may be used. Also, for example, in the fifth embodiment, a first trained model generation device 38, a second trained model generation device 39, and a prediction device 18 are used, but a device combining the functions of the first trained model generation device 38 and the second trained model generation device 39 and the prediction device 18 may be used, or a device combining the functions of the first trained model generation device 38, the second trained model generation device 39, and the prediction device 10 may be used.

[0296] (14) The functions of each device described in the above embodiment may be rearranged. For example, in the first embodiment, the prediction device 10 is described as having a function for predicting the probability of a fall and a function for determining whether intervention is necessary, and the model generation device 30 is described as having a function for generating the trained model 2100. However, instead of the prediction device 10 and the model generation device 30, a device may be provided that has a function for generating the trained model 2100 and a function for predicting the probability of a fall, and a device that has a function for determining whether intervention is necessary.

[0297] (15) In the above embodiment, data is collected from a subject to generate a trained model in the model generation phase, and the probability of a fall is predicted for the subject in the prediction phase. However, the subject may be the same as the subject. That is, the subject involved in generating the trained model in the model generation phase may be used as the subject to predict the probability of a fall in the prediction phase. Furthermore, information obtained from the subject to predict the probability of a fall in the prediction phase may be used to generate the trained model in the model generation phase.

[0298] (16) This is also related to (15) above, but the data acquired in the prediction phase may be used in the model generation phase. That is, each trained model may be updated as needed.

[0299] Note that, for convenience, the above description has been divided into machine learning models (e.g., machine learning using a neural network) and statistical method models (statistical models, e.g., models using a logistic function). However, because a statistical model is a model that shows the relationship between explanatory variables and a target variable using optimal parameters, it has in common with a trained model in which parameters have been optimized, in that it is a model indicated by optimized parameters. In other words, both machine learning models and statistical models (as well as calculation formulas) are models in which parameters optimized for prediction (estimation) are set. The parameters of a statistical model (e.g., a model using a logistic function) may be determined by machine learning, or the process of determining the parameters in a statistical model may be referred to as learning. In other words, a model in which parameters have been set may be referred to as a trained model. Furthermore, the model (model with set parameters) used for prediction (estimation) may be a model generated by a single method (e.g., any of neural network, support vector machine, random forest, gradient boosting, and logistic regression analysis), or may be a model generated by combining multiple methods (e.g., two or more of neural network, support vector machine, random forest, gradient boosting, and logistic regression analysis). In other words, prediction (estimation) may be performed using a model (model with set parameters) generated by a single method, or may be performed using a model generated by combining multiple methods.

[0300] The contents described in the above embodiment include at least the following configuration. Note that in the above embodiment, for convenience of explanation, the sensor information and the feature amount information are distinguished from each other. However, as described above, the feature amount information is information based on the sensor information, and both are information that is detected directly or indirectly. Therefore, the sensor information and the feature amount information may also be referred to as detected information.

[0301] (1) A correspondence information storage unit that stores correspondence information (e.g., trained models 2100, 2120, 2200, 2220, 2300, 2400, first trained models 2500, 2520, 2900, second trained models 2600, 2800, statistical models, conversion tables) in which detection information (e.g., sensor information or feature information based on the sensor information) obtained by sensing at least a part of the body is pre-associated with prediction information regarding future falls (e.g., probability of falling, estimated diagnosis information), and a correspondence information storage unit that stores correspondence information (e.g., trained models 2100, 2120, 2200, 2220, 2300, 2400, first trained models 2500, 2520, 2900, second trained models 2600, 2800, statistical models, conversion tables). A prediction system comprising: a prediction unit (e.g., fall prediction unit 102 of prediction device 10, fall prediction unit 122 of prediction device 12, fall prediction unit 142 of prediction device 14, fall prediction unit of prediction device 16, diagnostic value estimation unit 182 and fall prediction unit 183 of prediction device 18, diagnostic value estimation unit 202 and fall prediction unit 203 of prediction device 20, diagnostic value estimation unit 222 and fall prediction unit 223 of prediction device 22) that predicts information regarding a future fall of the subject based on the corresponding information when the detection information of the subject is input.

[0302] (2) The prediction system according to (1), wherein at least one part of the body is a leg.

[0303] (3) The prediction information is fall prediction information (e.g., fall probability) that indicates the possibility of a fall in the future, and the prediction unit (e.g., fall prediction unit 102 of prediction device 10, fall prediction unit 122 of prediction device 12, fall prediction unit 142 of prediction device 14, fall prediction unit 16, fall prediction unit 183 of prediction device 18, fall prediction unit 203 of prediction device 20, fall prediction unit 223 of prediction device 22) predicts the possibility of the subject falling in the future, in the prediction system described in (1) or (2).

[0304] (4) The correspondence information includes first correspondence information (e.g., first trained models 2500, 2520, 2900, statistical model, conversion table) in which the detection information is pre-associated with gait evaluation information (e.g., diagnostic information) representing an evaluation of gait abnormalities, and second correspondence information (e.g., second trained models 2600, 2800, statistical model, conversion table) in which the gait evaluation information is pre-associated with the prediction information, and the prediction unit (the diagnostic value estimation unit 182 and the fall prediction unit 183 of the prediction device 18, the diagnostic value estimation unit 202 and the fall prediction unit 203 of the prediction device 20, the diagnostic value estimation unit 222 and the fall prediction unit 223 of the prediction device 22) estimates the gait evaluation information based on the first correspondence information when the detection information of the subject is input, and predicts information regarding the subject's future fall based on the estimated gait evaluation information and the second correspondence information.

[0305] (5) A history information acquisition unit (for example, a configuration in which the model generation devices 30, 32, 34, and 36 acquire an input data set including the fall history information) that acquires, for a plurality of subjects, fall history information indicating whether or not a fall occurred within a specific period after the detection of the detection information, and a learning information generation unit (for example, a configuration in which the feature extraction unit 301 of the model generation device 30 and the model generation device 36 acquire an input data set including the fall history information) that generates, for the plurality of subjects, learning information (learning data sets 1100, 1120, 1200, 1220, 1300, 1400, and 1500) in which the detection information and the fall history information are associated with each other. a feature extraction unit 321 of the model generation device 32, a feature extraction unit of the model generation device 34, and a feature extraction unit of the model generation device 36), and a learning unit (e.g., the model generation unit 302 of the model generation device 30, the model generation unit 322 of the model generation device 32, the model generation unit 342 of the model generation device 34, and the model generation unit of the model generation device 36) that generates the correspondence information (e.g., trained models 2100, 2120, 2200, 2220, 2300, 2400, statistical model, conversion table) based on the learning information.

[0306] (6) A prediction system according to any one of (1) to (5), wherein the detected information is bending information representing at least one of flexion and extension of the foot during walking, pressure information representing sole pressure, or acceleration information representing foot acceleration.

[0307] (7) The correspondence information (e.g., trained model 2300, 2400, statistical model, conversion table) is further pre-associated with at least one attribute information (e.g., part of questionnaire information) of attribute information including age or gender, medical history information, information on experience of bone, joint, or muscle abnormalities, nursing care service information, medication information, or exercise habit information, and the prediction unit predicts the possibility of the subject falling in the future based on the correspondence information when the attribute information and the detection information of the subject are input.

[0308] (8) The corresponding information is further pre-associated with past fall history information (e.g., questionnaire information including as an item whether or not a fall occurred within the last three months) indicating whether or not a fall occurred within a specific period before the detection of the detection information, and the prediction unit, when the past fall history information and the detection information of the subject are input, predicts the possibility of the subject falling in the future based on the corresponding information.

[0309] (9) The corresponding information is further associated in advance with musculoskeletal evaluation information (e.g., part of questionnaire information) evaluated by answering multiple questions as a method of evaluating a person's functions of standing, walking, or sitting, which are related to the physical movements necessary for daily life, and the prediction unit, when the musculoskeletal evaluation information and the detection information of the subject are input, predicts the possibility of the subject falling in the future based on the corresponding information.

[0310] (10) A prediction system described in any of (1) to (9), in which the corresponding information is further pre-associated with motor function information (e.g., part of questionnaire information) that indicates subjective symptoms of decreased motor function, and the prediction unit, when the motor function information and the detection information of the subject are input, predicts the possibility of the subject falling in the future based on the corresponding information.

[0311] (11) A prediction system according to any one of (1) to (10), wherein the corresponding information is further pre-associated with obesity level information (part of the questionnaire information) calculated from height and weight, and the prediction unit, when the obesity level information and the detection information of the subject are input, predicts the possibility of the subject falling in the future based on the corresponding information.

[0312] (12) A prediction system according to any one of (1) to (11), wherein the correspondence information is a trained model trained using a neural network, a support vector machine, a random forest, gradient boosting, logistic regression analysis, or a combination of two or more of these techniques.

[0313] (13) A prediction system described in any one of (1) to (12), wherein the prediction information is intervention necessity information indicating whether intervention is necessary to deal with future falls (for example, information determining whether expert intervention is necessary based on the probability of a fall, etc.), and the prediction unit (for example, an intervention necessity determination unit (or fall prediction unit) of each prediction device) predicts whether intervention is necessary to deal with future falls.

[0314] (14) The correspondence information includes third correspondence information (e.g., learned models 2100, 2200, 2300, 2400, statistical model, conversion table) in which the detection information (e.g., sensor information or feature information based on the sensor information) is pre-associated with fall prediction information (e.g., fall probability) that indicates the possibility of a fall in the future, and fourth correspondence information in which the fall prediction information (e.g., fall probability) is pre-associated with the intervention necessity information (e.g., a table in which, when the threshold for fall probability is 60%, a value of fall probability less than 60% is associated with the fact that no intervention is required, and a value of fall probability 60% or more is associated with the fact that intervention is required), and the prediction unit, when the detection information of the subject is input, estimates the fall prediction information based on the third correspondence information and predicts the necessity of intervention for the subject to deal with a future fall based on the estimated fall prediction information and the fourth correspondence information.

[0315] (15) A prediction system according to (13) or (14), comprising: an evaluation information acquisition unit that acquires the intervention necessity information (e.g., information evaluated based on multiple diagnostic information, multiple questionnaire information, etc.) evaluated by an evaluator with reference to the walking state of multiple subjects within a specific period after the detection information is detected; a learning information generation unit that generates learning information in which the detection information and the intervention necessity information are associated with each other for the multiple subjects (e.g., a learning dataset including the detection information of each subject as an input sample and the intervention necessity information of each subject as an output sample); and a learning unit that generates the correspondence information based on the learning information (e.g., a trained model that outputs a judgment result on the necessity of intervention for the subject from an output layer when sensor information (or feature information) of the subject is input to an input layer).

[0316] (16) A prediction system according to any one of (1) to (15), wherein the prediction unit is provided with an output unit (e.g., an output unit of each prediction device) that predicts the possibility of the subject falling in the future, and outputs output information according to whether intervention is necessary to deal with future falls and a classification based on the possibility of falling in the future.

[0317] (17) A prediction method in a prediction system, comprising: a correspondence information storage process in which a correspondence information storage unit stores correspondence information in which detection information obtained by sensing at least a part of the body is pre-associated with prediction information regarding future falls; and a prediction process in which a prediction unit, when the detection information of a subject is input, predicts information regarding the subject's future falls based on the correspondence information.

[0318] (18) A prediction device comprising: a correspondence information reference unit that references correspondence information in which detection information sensed from at least a part of the body is pre-associated with predicted information regarding future falls; and a prediction unit that, when the detection information of a subject is input, predicts information regarding the subject's future falls based on the correspondence information.

[0319] (19) A learning device comprising: a learning information acquisition unit that acquires learning information in which detection information obtained by sensing at least a part of the body and prediction information regarding future falls are associated for each of a plurality of subjects; and a learning unit that performs machine learning based on the learning information.

[0320] (20) A program for causing a computer to execute a correspondence information reference means for referencing correspondence information in which detection information sensed on at least a part of the body is pre-associated with predicted information regarding future falls, and a prediction means for predicting information regarding a future fall of a subject based on the correspondence information when the detection information of the subject is input.

[0321] (21) A program for causing a computer to execute a learning information acquisition means for acquiring learning information in which detection information obtained by sensing at least a part of the body and prediction information regarding future falls are associated for each of a plurality of subjects, and a learning means for performing machine learning based on the learning information.

[0322] (22) A trained model in which detection information sensed from at least a part of the body and prediction information regarding future falls are generated by machine learning based on learning information associated with each of multiple subjects, and which predicts information regarding a subject's future falls when the detection information of the subject is input.

[0323] A program for realizing the functions of the devices according to the above-described embodiments (e.g., a prediction device, a model generation device, a first trained model generation device, a second trained model generation device, a training dataset generation device, etc.) may be stored in a computer-readable storage medium, and the program stored in this storage medium may be read into a computer system and executed to perform processing. Note that the term "computer system" here may also include hardware such as an operating system (OS) or peripheral devices. Furthermore, the term "computer-readable storage medium" refers to a storage device such as a flexible disk, a magneto-optical disk, a read-only memory (ROM), a writable non-volatile memory such as a flash memory, a portable medium such as a DVD (Digital Versatile Disc), or a hard disk built into a computer system.

[0324] Furthermore, the term "computer-readable storage medium" also includes a medium that stores a program for a certain period of time, such as a volatile memory (e.g., DRAM (Dynamic Random Access Memory)) within an information processing device or a client computer system when the program is transmitted via a network such as the Internet or a communication line such as a telephone line. The program may also be transmitted from a computer system storing the program in a storage device or the like to another computer system via a transmission medium or by transmission waves within the transmission medium. Here, the "transmission medium" that transmits the program refers to a medium that has the function of transmitting information, such as a network (communication network) such as the Internet or a communication line (communication line) such as a telephone line. The program may also be a so-called differential file (differential program) that can realize the aforementioned functions in combination with a program already stored in the computer system. [Explanation of symbols]

[0325] 10, 12, 14, 16, 18, 20, 22...Prediction device 30, 32, 34, 36...Model generation device 38, 42...First trained model generation device 39, 41...Second trained model generation device 1100, 1120, 1200, 1220, 1300, 1400, 1500, 1520... 1600, 1800, 1900... Training dataset 1110, 1210, 1510...input dataset 2100, 2120, 2200, 2220, 2300, 2400... Trained model 2500, 2520, 2900...First trained model 2600, 2800...Second trained model

Claims

1. a correspondence information storage unit that stores correspondence information in which detection information obtained by sensing at least a part of a body and prediction information regarding a future fall are previously associated with each other; a prediction unit that predicts information regarding a future fall of the subject based on the correspondence information when the detection information of the subject is input; A prediction system comprising:

2. The at least one body part is a leg. The prediction system of claim 1 .

3. The prediction information is fall prediction information indicating a possibility of a fall in the future, The prediction unit predicts a possibility that the subject will fall in the future. The prediction system of claim 1 .

4. The prediction information is fall prediction information indicating a possibility of a fall in the future, The prediction unit predicts a possibility that the subject will fall in the future. The prediction system of claim 2 .

5. The correspondence information includes first correspondence information in which the detection information and gait evaluation information representing an evaluation of gait abnormality are previously associated with each other, and second correspondence information in which the gait evaluation information and the prediction information are previously associated with each other, The prediction unit, when detection information of the subject is input, estimates the gait evaluation information based on the first correspondence information, and predicts information regarding a future fall of the subject based on the estimated gait evaluation information and the second correspondence information. A prediction system according to any one of claims 1 to 4.

6. a history information acquiring unit that acquires, for a plurality of subjects, fall history information indicating whether or not a fall occurred within a specific period after the detection information is detected; a learning information generating unit that generates learning information in which the detection information and the fall history information are associated with each other for the plurality of subjects; a learning unit that generates the correspondence information based on the learning information; The prediction system of claim 1 .

7. A history information acquisition unit that acquires fall history information indicating whether or not a fall occurred within a specific period after the detection information is detected for a plurality of subjects; a learning information generating unit that generates learning information in which the detection information and the fall history information are associated with each other for the plurality of subjects; a learning unit that generates the correspondence information based on the learning information; The prediction system of claim 2 .

8. A history information acquisition unit that acquires fall history information indicating whether or not a fall occurred within a specific period after the detection information is detected for a plurality of subjects; a learning information generating unit that generates learning information in which the detection information and the fall history information are associated with each other for the plurality of subjects; a learning unit that generates the correspondence information based on the learning information; The prediction system of claim 3 .

9. A history information acquisition unit that acquires fall history information indicating whether or not a fall occurred within a specific period after the detection information is detected for a plurality of subjects; a learning information generating unit that generates learning information in which the detection information and the fall history information are associated with each other for the plurality of subjects; a learning unit that generates the correspondence information based on the learning information; The prediction system of claim 4 .

10. A history information acquisition unit that acquires fall history information indicating whether or not a fall occurred within a specific period after the detection information is detected for a plurality of subjects; a learning information generating unit that generates learning information in which the detection information and the fall history information are associated with each other for the plurality of subjects; a learning unit that generates the correspondence information based on the learning information; The prediction system of claim 5 .

11. The detection information is bending information indicating at least one of bending and extension of the foot during walking, pressure information indicating the sole pressure, or acceleration information indicating the acceleration of the foot. A prediction system according to any one of claims 1 to 4 and claims 6 to 9.

12. The corresponding information is further associated in advance with at least one of attribute information including age or sex, medical history information, information on experience of abnormalities in bones, joints, or muscles, care service information, medication information, and exercise habit information; The prediction unit predicts a possibility that the subject will fall in the future based on the correspondence information when the attribute information and the detection information of the subject are input. A prediction system according to any one of claims 1 to 4 and claims 6 to 9.

13. The corresponding information is further previously associated with past fall history information indicating whether or not a fall occurred within a specific period of time before the detection of the detection information, The prediction unit predicts a possibility that the subject will fall in the future based on the correspondence information when the past fall history information and the detection information of the subject are input. A prediction system according to any one of claims 1 to 4 and claims 6 to 9.

14. The correspondence information is further associated in advance with musculoskeletal evaluation information evaluated by answering a plurality of questions as a method for evaluating a person's function of standing, walking, or sitting, which is related to the movement of the body necessary for daily life; The prediction unit predicts a possibility that the subject will fall in the future based on the correspondence information when the locomotor system evaluation information and the detection information of the subject are input. A prediction system according to any one of claims 1 to 4 and claims 6 to 9.

15. The correspondence information is further previously associated with motor function information indicating subjective symptoms of motor function decline, The prediction unit predicts a possibility that the subject will fall in the future based on the correspondence information when the motor function information and the detection information of the subject are input. A prediction system according to any one of claims 1 to 4 and claims 6 to 9.

16. The correspondence information is further previously associated with obesity level information calculated from height and weight, The prediction unit predicts a possibility that the subject will fall in the future based on the correspondence information when the obesity level information and the detection information of the subject are input. A prediction system according to any one of claims 1 to 4 and claims 6 to 9.

17. The correspondence information is a trained model trained using a neural network, a support vector machine, a random forest, gradient boosting, logistic regression analysis, or a combination of two or more of these techniques. A prediction system according to any one of claims 1 to 4 and claims 6 to 9.

18. The prediction information is intervention necessity information indicating whether or not intervention is necessary to deal with a future fall, The prediction unit predicts whether or not intervention is required to address a future fall. A prediction system according to any one of claims 1 to 4 and claims 6 to 9.

19. The correspondence information includes third correspondence information in which the detection information is previously associated with fall prediction information indicating a possibility of a fall in the future, and fourth correspondence information in which the fall prediction information is previously associated with the intervention necessity information, The prediction unit, when the detection information of the subject is input, estimates the fall prediction information based on the third correspondence information, and predicts whether or not intervention is necessary to deal with a future fall of the subject based on the estimated fall prediction information and the fourth correspondence information.

20. The prediction system of claim 18.

20. an evaluation information acquisition unit that acquires the intervention necessity information evaluated by an evaluator with reference to a walking state within a specific period after the detection information is detected for a plurality of subjects; a learning information generating unit that generates learning information in which the detection information and the intervention necessity information are associated with each other for the plurality of subjects; a learning unit that generates the correspondence information based on the learning information; 20. The prediction system of claim 19, comprising:

21. The prediction unit predicts a possibility that the subject will fall in the future, An output unit that outputs output information according to the necessity of intervention to deal with a future fall and the classification based on the likelihood of the future fall. A prediction system according to any one of claims 1 to 4 and claims 6 to 9.

22. The prediction unit predicts a possibility that the subject will fall in the future, An output unit that outputs output information according to the necessity of intervention to deal with a future fall and the classification based on the likelihood of the future fall.

20. The prediction system of claim 19.

23. A prediction method in a prediction system, comprising: a correspondence information storage step in which the correspondence information storage unit stores correspondence information in which detection information obtained by sensing at least a part of the body is associated in advance with prediction information regarding a future fall; a prediction step of predicting information regarding a future fall of the subject based on the correspondence information when the detection information of the subject is input by a prediction unit; A prediction method having the following structure:

24. a correspondence information reference unit that references correspondence information in which detection information obtained by sensing at least a part of a body and prediction information regarding a future fall are previously associated with each other; a prediction unit that predicts information regarding a future fall of the subject based on the correspondence information when the detection information of the subject is input; A prediction device comprising:

25. a learning information acquisition unit that acquires learning information in which detection information obtained by sensing at least a part of a body and prediction information regarding a future fall are associated with each of a plurality of subjects; and a learning unit that performs machine learning based on the learning information. A learning device comprising:

26. On the computer, a correspondence information reference means for referring to correspondence information in which detection information obtained by sensing at least a part of the body and prediction information regarding a future fall are previously associated with each other; a prediction means for predicting information regarding a future fall of the subject based on the corresponding information when the detection information of the subject is input; A program for executing the above.

27. On the computer, a learning information acquiring means for acquiring learning information in which detection information obtained by sensing at least a part of a body and prediction information regarding a future fall are associated with each of a plurality of subjects; A learning means for performing machine learning based on the learning information; A program for executing the above.

28. Detection information obtained by sensing at least a part of the body and prediction information regarding a future fall are generated by machine learning based on learning information associated with each of the multiple subjects; A trained model for predicting information regarding a subject's future falls when the detection information of the subject is input.

Images